i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001, 2002, 2003 Fabrice Bellard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This library 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 GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  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     TREG_EAX = 0,
  40     TREG_ECX,
  41     TREG_EDX,
  42     TREG_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 TREG_EAX /* single word int return register */
  54 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  55 #define REG_FRET TREG_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 /* maximum alignment (for aligned attribute support) */
  71 #define MAX_ALIGN     8
  72
  73 /******************************************************/
  74 /* ELF defines */
  75
  76 #define EM_TCC_TARGET EM_386
  77
  78 /* relocation type for 32 bit data relocation */
  79 #define R_DATA_32   R_386_32
  80 #define R_JMP_SLOT  R_386_JMP_SLOT
  81 #define R_COPY      R_386_COPY
  82
  83 #define ELF_START_ADDR 0x08048000
  84 #define ELF_PAGE_SIZE  0x1000
  85
  86 /******************************************************/
  87
  88 static unsigned long func_sub_sp_offset;
  89 static unsigned long func_bound_offset;
  90 static int func_ret_sub;
  91
  92 /* XXX: make it faster ? */
  93 void g(int c)
  94 {
  95     int ind1;
  96     ind1 = ind + 1;
  97     if (ind1 > cur_text_section->data_allocated)
  98         section_realloc(cur_text_section, ind1);
  99     cur_text_section->data[ind] = c;
 100     ind = ind1;
 101 }
 102
 103 void o(int c)
 104 {
 105     while (c) {
 106         g(c);
 107         c = c / 256;
 108     }
 109 }
 110
 111 void gen_le32(int c)
 112 {
 113     g(c);
 114     g(c >> 8);
 115     g(c >> 16);
 116     g(c >> 24);
 117 }
 118
 119 /* output a symbol and patch all calls to it */
 120 void gsym_addr(int t, int a)
 121 {
 122     int n, *ptr;
 123     while (t) {
 124         ptr = (int *)(cur_text_section->data + t);
 125         n = *ptr; /* next value */
 126         *ptr = 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 static int oad(int c, int s)
 142 {
 143     int ind1;
 144
 145     o(c);
 146     ind1 = ind + 4;
 147     if (ind1 > cur_text_section->data_allocated)
 148         section_realloc(cur_text_section, ind1);
 149     *(int *)(cur_text_section->data + ind) = s;
 150     s = ind;
 151     ind = ind1;
 152     return s;
 153 }
 154
 155 /* output constant with relocation if 'r & VT_SYM' is true */
 156 static void gen_addr32(int r, Sym *sym, int c)
 157 {
 158     if (r & VT_SYM)
 159         greloc(cur_text_section, sym, ind, R_386_32);
 160     gen_le32(c);
 161 }
 162
 163 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 164    opcode bits */
 165 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 166 {
 167     op_reg = op_reg << 3;
 168     if ((r & VT_VALMASK) == VT_CONST) {
 169         /* constant memory reference */
 170         o(0x05 | op_reg);
 171         gen_addr32(r, sym, c);
 172     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 173         /* currently, we use only ebp as base */
 174         if (c == (char)c) {
 175             /* short reference */
 176             o(0x45 | op_reg);
 177             g(c);
 178         } else {
 179             oad(0x85 | op_reg, c);
 180         }
 181     } else {
 182         g(0x00 | op_reg | (r & VT_VALMASK));
 183     }
 184 }
 185
 186
 187 /* load 'r' from value 'sv' */
 188 void load(int r, SValue *sv)
 189 {
 190     int v, t, ft, fc, fr;
 191     SValue v1;
 192
 193     fr = sv->r;
 194     ft = sv->type.t;
 195     fc = sv->c.ul;
 196
 197     v = fr & VT_VALMASK;
 198     if (fr & VT_LVAL) {
 199         if (v == VT_LLOCAL) {
 200             v1.type.t = VT_INT;
 201             v1.r = VT_LOCAL | VT_LVAL;
 202             v1.c.ul = fc;
 203             load(r, &v1);
 204             fr = r;
 205         }
 206         if ((ft & VT_BTYPE) == VT_FLOAT) {
 207             o(0xd9); /* flds */
 208             r = 0;
 209         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 210             o(0xdd); /* fldl */
 211             r = 0;
 212         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 213             o(0xdb); /* fldt */
 214             r = 5;
 215         } else if ((ft & VT_TYPE) == VT_BYTE) {
 216             o(0xbe0f);   /* movsbl */
 217         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 218             o(0xb60f);   /* movzbl */
 219         } else if ((ft & VT_TYPE) == VT_SHORT) {
 220             o(0xbf0f);   /* movswl */
 221         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 222             o(0xb70f);   /* movzwl */
 223         } else {
 224             o(0x8b);     /* movl */
 225         }
 226         gen_modrm(r, fr, sv->sym, fc);
 227     } else {
 228         if (v == VT_CONST) {
 229             o(0xb8 + r); /* mov $xx, r */
 230             gen_addr32(fr, sv->sym, fc);
 231         } else if (v == VT_LOCAL) {
 232             o(0x8d); /* lea xxx(%ebp), r */
 233             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 234         } else if (v == VT_CMP) {
 235             oad(0xb8 + r, 0); /* mov $0, r */
 236             o(0x0f); /* setxx %br */
 237             o(fc);
 238             o(0xc0 + r);
 239         } else if (v == VT_JMP || v == VT_JMPI) {
 240             t = v & 1;
 241             oad(0xb8 + r, t); /* mov $1, r */
 242             o(0x05eb); /* jmp after */
 243             gsym(fc);
 244             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 245         } else if (v != r) {
 246             o(0x89);
 247             o(0xc0 + r + v * 8); /* mov v, r */
 248         }
 249     }
 250 }
 251
 252 /* store register 'r' in lvalue 'v' */
 253 void store(int r, SValue *v)
 254 {
 255     int fr, bt, ft, fc;
 256
 257     ft = v->type.t;
 258     fc = v->c.ul;
 259     fr = v->r & VT_VALMASK;
 260     bt = ft & VT_BTYPE;
 261     /* XXX: incorrect if float reg to reg */
 262     if (bt == VT_FLOAT) {
 263         o(0xd9); /* fsts */
 264         r = 2;
 265     } else if (bt == VT_DOUBLE) {
 266         o(0xdd); /* fstpl */
 267         r = 2;
 268     } else if (bt == VT_LDOUBLE) {
 269         o(0xc0d9); /* fld %st(0) */
 270         o(0xdb); /* fstpt */
 271         r = 7;
 272     } else {
 273         if (bt == VT_SHORT)
 274             o(0x66);
 275         if (bt == VT_BYTE)
 276             o(0x88);
 277         else
 278             o(0x89);
 279     }
 280     if (fr == VT_CONST ||
 281         fr == VT_LOCAL ||
 282         (v->r & VT_LVAL)) {
 283         gen_modrm(r, v->r, v->sym, fc);
 284     } else if (fr != r) {
 285         o(0xc0 + fr + r * 8); /* mov r, fr */
 286     }
 287 }
 288
 289 static void gadd_sp(int val)
 290 {
 291     if (val == (char)val) {
 292         o(0xc483);
 293         g(val);
 294     } else {
 295         oad(0xc481, val); /* add $xxx, %esp */
 296     }
 297 }
 298
 299 /* 'is_jmp' is '1' if it is a jump */
 300 static void gcall_or_jmp(int is_jmp)
 301 {
 302     int r;
 303     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 304         /* constant case */
 305         if (vtop->r & VT_SYM) {
 306             /* relocation case */
 307             greloc(cur_text_section, vtop->sym,
 308                    ind + 1, R_386_PC32);
 309         } else {
 310             /* put an empty PC32 relocation */
 311             put_elf_reloc(symtab_section, cur_text_section,
 312                           ind + 1, R_386_PC32, 0);
 313         }
 314         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 315     } else {
 316         /* otherwise, indirect call */
 317         r = gv(RC_INT);
 318         o(0xff); /* call/jmp *r */
 319         o(0xd0 + r + (is_jmp << 4));
 320     }
 321 }
 322
 323 /* Generate function call. The function address is pushed first, then
 324    all the parameters in call order. This functions pops all the
 325    parameters and the function address. */
 326 void gfunc_call(int nb_args)
 327 {
 328     int size, align, r, args_size, i;
 329     Sym *func_sym;
 330
 331     args_size = 0;
 332     for(i = 0;i < nb_args; i++) {
 333         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 334             size = type_size(&vtop->type, &align);
 335             /* align to stack align size */
 336             size = (size + 3) & ~3;
 337             /* allocate the necessary size on stack */
 338             oad(0xec81, size); /* sub $xxx, %esp */
 339             /* generate structure store */
 340             r = get_reg(RC_INT);
 341             o(0x89); /* mov %esp, r */
 342             o(0xe0 + r);
 343             vset(&vtop->type, r | VT_LVAL, 0);
 344             vswap();
 345             vstore();
 346             args_size += size;
 347         } else if (is_float(vtop->type.t)) {
 348             gv(RC_FLOAT); /* only one float register */
 349             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 350                 size = 4;
 351             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 352                 size = 8;
 353             else
 354                 size = 12;
 355             oad(0xec81, size); /* sub $xxx, %esp */
 356             if (size == 12)
 357                 o(0x7cdb);
 358             else
 359                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 360             g(0x24);
 361             g(0x00);
 362             args_size += size;
 363         } else {
 364             /* simple type (currently always same size) */
 365             /* XXX: implicit cast ? */
 366             r = gv(RC_INT);
 367             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 368                 size = 8;
 369                 o(0x50 + vtop->r2); /* push r */
 370             } else {
 371                 size = 4;
 372             }
 373             o(0x50 + r); /* push r */
 374             args_size += size;
 375         }
 376         vtop--;
 377     }
 378     save_regs(0); /* save used temporary registers */
 379     func_sym = vtop->type.ref;
 380     gcall_or_jmp(0);
 381     if (args_size && func_sym->r == FUNC_CDECL)
 382         gadd_sp(args_size);
 383     vtop--;
 384 }
 385
 386 /* generate function prolog of type 't' */
 387 void gfunc_prolog(CType *func_type)
 388 {
 389     int addr, align, size, func_call;
 390     Sym *sym;
 391     CType *type;
 392
 393     sym = func_type->ref;
 394     func_call = sym->r;
 395     addr = 8;
 396     /* if the function returns a structure, then add an
 397        implicit pointer parameter */
 398     func_vt = sym->type;
 399     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 400         func_vc = addr;
 401         addr += 4;
 402     }
 403     /* define parameters */
 404     while ((sym = sym->next) != NULL) {
 405         type = &sym->type;
 406         sym_push(sym->v & ~SYM_FIELD, type,
 407                  VT_LOCAL | VT_LVAL, addr);
 408         size = type_size(type, &align);
 409         size = (size + 3) & ~3;
 410 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 411         /* structs are passed as pointer */
 412         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 413             size = 4;
 414         }
 415 #endif
 416         addr += size;
 417     }
 418     func_ret_sub = 0;
 419     /* pascal type call ? */
 420     if (func_call == FUNC_STDCALL)
 421         func_ret_sub = addr - 8;
 422     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 423     func_sub_sp_offset = oad(0xec81, 0); /* sub $xxx, %esp */
 424     /* leave some room for bound checking code */
 425     if (do_bounds_check) {
 426         oad(0xb8, 0); /* lbound section pointer */
 427         oad(0xb8, 0); /* call to function */
 428         func_bound_offset = lbounds_section->data_offset;
 429     }
 430 }
 431
 432 /* generate function epilog */
 433 void gfunc_epilog(void)
 434 {
 435 #ifdef CONFIG_TCC_BCHECK
 436     if (do_bounds_check && func_bound_offset != lbounds_section->data_offset) {
 437         int saved_ind;
 438         int *bounds_ptr;
 439         Sym *sym, *sym_data;
 440         /* add end of table info */
 441         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 442         *bounds_ptr = 0;
 443         /* generate bound local allocation */
 444         saved_ind = ind;
 445         ind = func_sub_sp_offset + 4;
 446         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 447                                func_bound_offset, lbounds_section->data_offset);
 448         greloc(cur_text_section, sym_data,
 449                ind + 1, R_386_32);
 450         oad(0xb8, 0); /* mov %eax, xxx */
 451         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 452         greloc(cur_text_section, sym,
 453                ind + 1, R_386_PC32);
 454         oad(0xe8, -4);
 455         ind = saved_ind;
 456         /* generate bound check local freeing */
 457         o(0x5250); /* save returned value, if any */
 458         greloc(cur_text_section, sym_data,
 459                ind + 1, R_386_32);
 460         oad(0xb8, 0); /* mov %eax, xxx */
 461         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 462         greloc(cur_text_section, sym,
 463                ind + 1, R_386_PC32);
 464         oad(0xe8, -4);
 465         o(0x585a); /* restore returned value, if any */
 466     }
 467 #endif
 468     o(0xc9); /* leave */
 469     if (func_ret_sub == 0) {
 470         o(0xc3); /* ret */
 471     } else {
 472         o(0xc2); /* ret n */
 473         g(func_ret_sub);
 474         g(func_ret_sub >> 8);
 475     }
 476     /* align local size to word & save local variables */
 477     *(int *)(cur_text_section->data + func_sub_sp_offset) = (-loc + 3) & -4;
 478 }
 479
 480 /* generate a jump to a label */
 481 int gjmp(int t)
 482 {
 483     return psym(0xe9, t);
 484 }
 485
 486 /* generate a jump to a fixed address */
 487 void gjmp_addr(int a)
 488 {
 489     int r;
 490     r = a - ind - 2;
 491     if (r == (char)r) {
 492         g(0xeb);
 493         g(r);
 494     } else {
 495         oad(0xe9, a - ind - 5);
 496     }
 497 }
 498
 499 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 500 int gtst(int inv, int t)
 501 {
 502     int v, *p;
 503
 504     v = vtop->r & VT_VALMASK;
 505     if (v == VT_CMP) {
 506         /* fast case : can jump directly since flags are set */
 507         g(0x0f);
 508         t = psym((vtop->c.i - 16) ^ inv, t);
 509     } else if (v == VT_JMP || v == VT_JMPI) {
 510         /* && or || optimization */
 511         if ((v & 1) == inv) {
 512             /* insert vtop->c jump list in t */
 513             p = &vtop->c.i;
 514             while (*p != 0)
 515                 p = (int *)(cur_text_section->data + *p);
 516             *p = t;
 517             t = vtop->c.i;
 518         } else {
 519             t = gjmp(t);
 520             gsym(vtop->c.i);
 521         }
 522     } else {
 523         if (is_float(vtop->type.t)) {
 524             vpushi(0);
 525             gen_op(TOK_NE);
 526         }
 527         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 528             /* constant jmp optimization */
 529             if ((vtop->c.i != 0) != inv)
 530                 t = gjmp(t);
 531         } else {
 532             v = gv(RC_INT);
 533             o(0x85);
 534             o(0xc0 + v * 9);
 535             g(0x0f);
 536             t = psym(0x85 ^ inv, t);
 537         }
 538     }
 539     vtop--;
 540     return t;
 541 }
 542
 543 /* generate an integer binary operation */
 544 void gen_opi(int op)
 545 {
 546     int r, fr, opc, c;
 547
 548     switch(op) {
 549     case '+':
 550     case TOK_ADDC1: /* add with carry generation */
 551         opc = 0;
 552     gen_op8:
 553         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 554             /* constant case */
 555             vswap();
 556             r = gv(RC_INT);
 557             vswap();
 558             c = vtop->c.i;
 559             if (c == (char)c) {
 560                 /* XXX: generate inc and dec for smaller code ? */
 561                 o(0x83);
 562                 o(0xc0 | (opc << 3) | r);
 563                 g(c);
 564             } else {
 565                 o(0x81);
 566                 oad(0xc0 | (opc << 3) | r, c);
 567             }
 568         } else {
 569             gv2(RC_INT, RC_INT);
 570             r = vtop[-1].r;
 571             fr = vtop[0].r;
 572             o((opc << 3) | 0x01);
 573             o(0xc0 + r + fr * 8);
 574         }
 575         vtop--;
 576         if (op >= TOK_ULT && op <= TOK_GT) {
 577             vtop->r = VT_CMP;
 578             vtop->c.i = op;
 579         }
 580         break;
 581     case '-':
 582     case TOK_SUBC1: /* sub with carry generation */
 583         opc = 5;
 584         goto gen_op8;
 585     case TOK_ADDC2: /* add with carry use */
 586         opc = 2;
 587         goto gen_op8;
 588     case TOK_SUBC2: /* sub with carry use */
 589         opc = 3;
 590         goto gen_op8;
 591     case '&':
 592         opc = 4;
 593         goto gen_op8;
 594     case '^':
 595         opc = 6;
 596         goto gen_op8;
 597     case '|':
 598         opc = 1;
 599         goto gen_op8;
 600     case '*':
 601         gv2(RC_INT, RC_INT);
 602         r = vtop[-1].r;
 603         fr = vtop[0].r;
 604         vtop--;
 605         o(0xaf0f); /* imul fr, r */
 606         o(0xc0 + fr + r * 8);
 607         break;
 608     case TOK_SHL:
 609         opc = 4;
 610         goto gen_shift;
 611     case TOK_SHR:
 612         opc = 5;
 613         goto gen_shift;
 614     case TOK_SAR:
 615         opc = 7;
 616     gen_shift:
 617         opc = 0xc0 | (opc << 3);
 618         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 619             /* constant case */
 620             vswap();
 621             r = gv(RC_INT);
 622             vswap();
 623             c = vtop->c.i & 0x1f;
 624             o(0xc1); /* shl/shr/sar $xxx, r */
 625             o(opc | r);
 626             g(c);
 627         } else {
 628             /* we generate the shift in ecx */
 629             gv2(RC_INT, RC_ECX);
 630             r = vtop[-1].r;
 631             o(0xd3); /* shl/shr/sar %cl, r */
 632             o(opc | r);
 633         }
 634         vtop--;
 635         break;
 636     case '/':
 637     case TOK_UDIV:
 638     case TOK_PDIV:
 639     case '%':
 640     case TOK_UMOD:
 641     case TOK_UMULL:
 642         /* first operand must be in eax */
 643         /* XXX: need better constraint for second operand */
 644         gv2(RC_EAX, RC_ECX);
 645         r = vtop[-1].r;
 646         fr = vtop[0].r;
 647         vtop--;
 648         save_reg(TREG_EDX);
 649         if (op == TOK_UMULL) {
 650             o(0xf7); /* mul fr */
 651             o(0xe0 + fr);
 652             vtop->r2 = TREG_EDX;
 653             r = TREG_EAX;
 654         } else {
 655             if (op == TOK_UDIV || op == TOK_UMOD) {
 656                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 657                 o(0xf0 + fr);
 658             } else {
 659                 o(0xf799); /* cltd, idiv fr, %eax */
 660                 o(0xf8 + fr);
 661             }
 662             if (op == '%' || op == TOK_UMOD)
 663                 r = TREG_EDX;
 664             else
 665                 r = TREG_EAX;
 666         }
 667         vtop->r = r;
 668         break;
 669     default:
 670         opc = 7;
 671         goto gen_op8;
 672     }
 673 }
 674
 675 /* generate a floating point operation 'v = t1 op t2' instruction. The
 676    two operands are guaranted to have the same floating point type */
 677 /* XXX: need to use ST1 too */
 678 void gen_opf(int op)
 679 {
 680     int a, ft, fc, swapped, r;
 681
 682     /* convert constants to memory references */
 683     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 684         vswap();
 685         gv(RC_FLOAT);
 686         vswap();
 687     }
 688     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 689         gv(RC_FLOAT);
 690
 691     /* must put at least one value in the floating point register */
 692     if ((vtop[-1].r & VT_LVAL) &&
 693         (vtop[0].r & VT_LVAL)) {
 694         vswap();
 695         gv(RC_FLOAT);
 696         vswap();
 697     }
 698     swapped = 0;
 699     /* swap the stack if needed so that t1 is the register and t2 is
 700        the memory reference */
 701     if (vtop[-1].r & VT_LVAL) {
 702         vswap();
 703         swapped = 1;
 704     }
 705     if (op >= TOK_ULT && op <= TOK_GT) {
 706         /* load on stack second operand */
 707         load(TREG_ST0, vtop);
 708         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 709         if (op == TOK_GE || op == TOK_GT)
 710             swapped = !swapped;
 711         else if (op == TOK_EQ || op == TOK_NE)
 712             swapped = 0;
 713         if (swapped)
 714             o(0xc9d9); /* fxch %st(1) */
 715         o(0xe9da); /* fucompp */
 716         o(0xe0df); /* fnstsw %ax */
 717         if (op == TOK_EQ) {
 718             o(0x45e480); /* and $0x45, %ah */
 719             o(0x40fC80); /* cmp $0x40, %ah */
 720         } else if (op == TOK_NE) {
 721             o(0x45e480); /* and $0x45, %ah */
 722             o(0x40f480); /* xor $0x40, %ah */
 723             op = TOK_NE;
 724         } else if (op == TOK_GE || op == TOK_LE) {
 725             o(0x05c4f6); /* test $0x05, %ah */
 726             op = TOK_EQ;
 727         } else {
 728             o(0x45c4f6); /* test $0x45, %ah */
 729             op = TOK_EQ;
 730         }
 731         vtop--;
 732         vtop->r = VT_CMP;
 733         vtop->c.i = op;
 734     } else {
 735         /* no memory reference possible for long double operations */
 736         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 737             load(TREG_ST0, vtop);
 738             swapped = !swapped;
 739         }
 740
 741         switch(op) {
 742         default:
 743         case '+':
 744             a = 0;
 745             break;
 746         case '-':
 747             a = 4;
 748             if (swapped)
 749                 a++;
 750             break;
 751         case '*':
 752             a = 1;
 753             break;
 754         case '/':
 755             a = 6;
 756             if (swapped)
 757                 a++;
 758             break;
 759         }
 760         ft = vtop->type.t;
 761         fc = vtop->c.ul;
 762         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 763             o(0xde); /* fxxxp %st, %st(1) */
 764             o(0xc1 + (a << 3));
 765         } else {
 766             /* if saved lvalue, then we must reload it */
 767             r = vtop->r;
 768             if ((r & VT_VALMASK) == VT_LLOCAL) {
 769                 SValue v1;
 770                 r = get_reg(RC_INT);
 771                 v1.type.t = VT_INT;
 772                 v1.r = VT_LOCAL | VT_LVAL;
 773                 v1.c.ul = fc;
 774                 load(r, &v1);
 775                 fc = 0;
 776             }
 777
 778             if ((ft & VT_BTYPE) == VT_DOUBLE)
 779                 o(0xdc);
 780             else
 781                 o(0xd8);
 782             gen_modrm(a, r, vtop->sym, fc);
 783         }
 784         vtop--;
 785     }
 786 }
 787
 788 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 789    and 'long long' cases. */
 790 void gen_cvt_itof(int t)
 791 {
 792     save_reg(TREG_ST0);
 793     gv(RC_INT);
 794     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 795         /* signed long long to float/double/long double (unsigned case
 796            is handled generically) */
 797         o(0x50 + vtop->r2); /* push r2 */
 798         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 799         o(0x242cdf); /* fildll (%esp) */
 800         o(0x08c483); /* add $8, %esp */
 801     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 802                (VT_INT | VT_UNSIGNED)) {
 803         /* unsigned int to float/double/long double */
 804         o(0x6a); /* push $0 */
 805         g(0x00);
 806         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 807         o(0x242cdf); /* fildll (%esp) */
 808         o(0x08c483); /* add $8, %esp */
 809     } else {
 810         /* int to float/double/long double */
 811         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 812         o(0x2404db); /* fildl (%esp) */
 813         o(0x04c483); /* add $4, %esp */
 814     }
 815     vtop->r = TREG_ST0;
 816 }
 817
 818 /* convert fp to int 't' type */
 819 /* XXX: handle long long case */
 820 void gen_cvt_ftoi(int t)
 821 {
 822     int r, r2, size;
 823     Sym *sym;
 824     CType ushort_type;
 825
 826     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 827
 828     gv(RC_FLOAT);
 829     if (t != VT_INT)
 830         size = 8;
 831     else
 832         size = 4;
 833
 834     o(0x2dd9); /* ldcw xxx */
 835     sym = external_global_sym(TOK___tcc_int_fpu_control,
 836                               &ushort_type, VT_LVAL);
 837     greloc(cur_text_section, sym,
 838            ind, R_386_32);
 839     gen_le32(0);
 840
 841     oad(0xec81, size); /* sub $xxx, %esp */
 842     if (size == 4)
 843         o(0x1cdb); /* fistpl */
 844     else
 845         o(0x3cdf); /* fistpll */
 846     o(0x24);
 847     o(0x2dd9); /* ldcw xxx */
 848     sym = external_global_sym(TOK___tcc_fpu_control,
 849                               &ushort_type, VT_LVAL);
 850     greloc(cur_text_section, sym,
 851            ind, R_386_32);
 852     gen_le32(0);
 853
 854     r = get_reg(RC_INT);
 855     o(0x58 + r); /* pop r */
 856     if (size == 8) {
 857         if (t == VT_LLONG) {
 858             vtop->r = r; /* mark reg as used */
 859             r2 = get_reg(RC_INT);
 860             o(0x58 + r2); /* pop r2 */
 861             vtop->r2 = r2;
 862         } else {
 863             o(0x04c483); /* add $4, %esp */
 864         }
 865     }
 866     vtop->r = r;
 867 }
 868
 869 /* convert from one floating point type to another */
 870 void gen_cvt_ftof(int t)
 871 {
 872     /* all we have to do on i386 is to put the float in a register */
 873     gv(RC_FLOAT);
 874 }
 875
 876 /* computed goto support */
 877 void ggoto(void)
 878 {
 879     gcall_or_jmp(1);
 880     vtop--;
 881 }
 882
 883 /* bound check support functions */
 884 #ifdef CONFIG_TCC_BCHECK
 885
 886 /* generate a bounded pointer addition */
 887 void gen_bounded_ptr_add(void)
 888 {
 889     Sym *sym;
 890
 891     /* prepare fast i386 function call (args in eax and edx) */
 892     gv2(RC_EAX, RC_EDX);
 893     /* save all temporary registers */
 894     vtop -= 2;
 895     save_regs(0);
 896     /* do a fast function call */
 897     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
 898     greloc(cur_text_section, sym,
 899            ind + 1, R_386_PC32);
 900     oad(0xe8, -4);
 901     /* returned pointer is in eax */
 902     vtop++;
 903     vtop->r = TREG_EAX | VT_BOUNDED;
 904     /* address of bounding function call point */
 905     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 906 }
 907
 908 /* patch pointer addition in vtop so that pointer dereferencing is
 909    also tested */
 910 void gen_bounded_ptr_deref(void)
 911 {
 912     int func;
 913     int size, align;
 914     Elf32_Rel *rel;
 915     Sym *sym;
 916
 917     size = 0;
 918     /* XXX: put that code in generic part of tcc */
 919     if (!is_float(vtop->type.t)) {
 920         if (vtop->r & VT_LVAL_BYTE)
 921             size = 1;
 922         else if (vtop->r & VT_LVAL_SHORT)
 923             size = 2;
 924     }
 925     if (!size)
 926         size = type_size(&vtop->type, &align);
 927     switch(size) {
 928     case  1: func = TOK___bound_ptr_indir1; break;
 929     case  2: func = TOK___bound_ptr_indir2; break;
 930     case  4: func = TOK___bound_ptr_indir4; break;
 931     case  8: func = TOK___bound_ptr_indir8; break;
 932     case 12: func = TOK___bound_ptr_indir12; break;
 933     case 16: func = TOK___bound_ptr_indir16; break;
 934     default:
 935         error("unhandled size when derefencing bounded pointer");
 936         func = 0;
 937         break;
 938     }
 939
 940     /* patch relocation */
 941     /* XXX: find a better solution ? */
 942     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
 943     sym = external_global_sym(func, &func_old_type, 0);
 944     if (!sym->c)
 945         put_extern_sym(sym, NULL, 0, 0);
 946     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
 947 }
 948 #endif
 949
 950 /* end of X86 code generator */
 951 /*************************************************************/
 952