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 unsigned long func_sub_sp_offset;
  83 static unsigned long func_bound_offset;
  84 static int func_ret_sub;
  85
  86 void g(int c)
  87 {
  88     cur_text_section->data[ind] = c;
  89     ind++;
  90 }
  91
  92 void o(int c)
  93 {
  94     while (c) {
  95         g(c);
  96         c = c / 256;
  97     }
  98 }
  99
 100 void gen_le32(int c)
 101 {
 102     g(c);
 103     g(c >> 8);
 104     g(c >> 16);
 105     g(c >> 24);
 106 }
 107
 108 /* output a symbol and patch all calls to it */
 109 void gsym_addr(int t, int a)
 110 {
 111     int n, *ptr;
 112     while (t) {
 113         ptr = (int *)(cur_text_section->data + t);
 114         n = *ptr; /* next value */
 115         *ptr = a - t - 4;
 116         t = n;
 117     }
 118 }
 119
 120 void gsym(int t)
 121 {
 122     gsym_addr(t, ind);
 123 }
 124
 125 /* psym is used to put an instruction with a data field which is a
 126    reference to a symbol. It is in fact the same as oad ! */
 127 #define psym oad
 128
 129 /* instruction + 4 bytes data. Return the address of the data */
 130 int oad(int c, int s)
 131 {
 132     o(c);
 133     *(int *)(cur_text_section->data + ind) = s;
 134     s = ind;
 135     ind = ind + 4;
 136     return s;
 137 }
 138
 139 /* output constant with relocation if 'r & VT_SYM' is true */
 140 void gen_addr32(int r, int c)
 141 {
 142     if (!(r & VT_SYM)) {
 143         gen_le32(c);
 144     } else {
 145         greloc(cur_text_section,
 146                (Sym *)c, ind, R_386_32);
 147         gen_le32(0);
 148     }
 149 }
 150
 151 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 152    opcode bits */
 153 void gen_modrm(int op_reg, int r, int c)
 154 {
 155     op_reg = op_reg << 3;
 156     if ((r & VT_VALMASK) == VT_CONST) {
 157         /* constant memory reference */
 158         o(0x05 | op_reg);
 159         gen_addr32(r, c);
 160     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 161         /* currently, we use only ebp as base */
 162         if (c == (char)c) {
 163             /* short reference */
 164             o(0x45 | op_reg);
 165             g(c);
 166         } else {
 167             oad(0x85 | op_reg, c);
 168         }
 169     } else {
 170         g(0x00 | op_reg | (r & VT_VALMASK));
 171     }
 172 }
 173
 174
 175 /* load 'r' from value 'sv' */
 176 void load(int r, SValue *sv)
 177 {
 178     int v, t, ft, fc, fr;
 179     SValue v1;
 180
 181     fr = sv->r;
 182     ft = sv->t;
 183     fc = sv->c.ul;
 184
 185     v = fr & VT_VALMASK;
 186     if (fr & VT_LVAL) {
 187         if (v == VT_LLOCAL) {
 188             v1.t = VT_INT;
 189             v1.r = VT_LOCAL | VT_LVAL;
 190             v1.c.ul = fc;
 191             load(r, &v1);
 192             fr = r;
 193         }
 194         if ((ft & VT_BTYPE) == VT_FLOAT) {
 195             o(0xd9); /* flds */
 196             r = 0;
 197         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 198             o(0xdd); /* fldl */
 199             r = 0;
 200         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 201             o(0xdb); /* fldt */
 202             r = 5;
 203         } else if ((ft & VT_TYPE) == VT_BYTE) {
 204             o(0xbe0f);   /* movsbl */
 205         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 206             o(0xb60f);   /* movzbl */
 207         } else if ((ft & VT_TYPE) == VT_SHORT) {
 208             o(0xbf0f);   /* movswl */
 209         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 210             o(0xb70f);   /* movzwl */
 211         } else {
 212             o(0x8b);     /* movl */
 213         }
 214         gen_modrm(r, fr, fc);
 215     } else {
 216         if (v == VT_CONST) {
 217             o(0xb8 + r); /* mov $xx, r */
 218             gen_addr32(fr, fc);
 219         } else if (v == VT_LOCAL) {
 220             o(0x8d); /* lea xxx(%ebp), r */
 221             gen_modrm(r, VT_LOCAL, fc);
 222         } else if (v == VT_CMP) {
 223             oad(0xb8 + r, 0); /* mov $0, r */
 224             o(0x0f); /* setxx %br */
 225             o(fc);
 226             o(0xc0 + r);
 227         } else if (v == VT_JMP || v == VT_JMPI) {
 228             t = v & 1;
 229             oad(0xb8 + r, t); /* mov $1, r */
 230             oad(0xe9, 5); /* jmp after */
 231             gsym(fc);
 232             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 233         } else if (v != r) {
 234             o(0x89);
 235             o(0xc0 + r + v * 8); /* mov v, r */
 236         }
 237     }
 238 }
 239
 240 /* store register 'r' in lvalue 'v' */
 241 void store(int r, SValue *v)
 242 {
 243     int fr, bt, ft, fc;
 244
 245     ft = v->t;
 246     fc = v->c.ul;
 247     fr = v->r & VT_VALMASK;
 248     bt = ft & VT_BTYPE;
 249     /* XXX: incorrect if float reg to reg */
 250     if (bt == VT_FLOAT) {
 251         o(0xd9); /* fsts */
 252         r = 2;
 253     } else if (bt == VT_DOUBLE) {
 254         o(0xdd); /* fstpl */
 255         r = 2;
 256     } else if (bt == VT_LDOUBLE) {
 257         o(0xc0d9); /* fld %st(0) */
 258         o(0xdb); /* fstpt */
 259         r = 7;
 260     } else {
 261         if (bt == VT_SHORT)
 262             o(0x66);
 263         if (bt == VT_BYTE)
 264             o(0x88);
 265         else
 266             o(0x89);
 267     }
 268     if (fr == VT_CONST ||
 269         fr == VT_LOCAL ||
 270         (v->r & VT_LVAL)) {
 271         gen_modrm(r, v->r, fc);
 272     } else if (fr != r) {
 273         o(0xc0 + fr + r * 8); /* mov r, fr */
 274     }
 275 }
 276
 277 /* start function call and return function call context */
 278 void gfunc_start(GFuncContext *c, int func_call)
 279 {
 280     c->args_size = 0;
 281     c->func_call = func_call;
 282 }
 283
 284 /* push function parameter which is in (vtop->t, vtop->c). Stack entry
 285    is then popped. */
 286 void gfunc_param(GFuncContext *c)
 287 {
 288     int size, align, r;
 289
 290     if ((vtop->t & VT_BTYPE) == VT_STRUCT) {
 291         size = type_size(vtop->t, &align);
 292         /* align to stack align size */
 293         size = (size + 3) & ~3;
 294         /* allocate the necessary size on stack */
 295         oad(0xec81, size); /* sub $xxx, %esp */
 296         /* generate structure store */
 297         r = get_reg(RC_INT);
 298         o(0x89); /* mov %esp, r */
 299         o(0xe0 + r);
 300         vset(vtop->t, r | VT_LVAL, 0);
 301         vswap();
 302         vstore();
 303         c->args_size += size;
 304     } else if (is_float(vtop->t)) {
 305         gv(RC_FLOAT); /* only one float register */
 306         if ((vtop->t & VT_BTYPE) == VT_FLOAT)
 307             size = 4;
 308         else if ((vtop->t & VT_BTYPE) == VT_DOUBLE)
 309             size = 8;
 310         else
 311             size = 12;
 312         oad(0xec81, size); /* sub $xxx, %esp */
 313         if (size == 12)
 314             o(0x7cdb);
 315         else
 316             o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 317         g(0x24);
 318         g(0x00);
 319         c->args_size += size;
 320     } else {
 321         /* simple type (currently always same size) */
 322         /* XXX: implicit cast ? */
 323         r = gv(RC_INT);
 324         if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 325             size = 8;
 326             o(0x50 + vtop->r2); /* push r */
 327         } else {
 328             size = 4;
 329         }
 330         o(0x50 + r); /* push r */
 331         c->args_size += size;
 332     }
 333     vtop--;
 334 }
 335
 336 static void gadd_sp(int val)
 337 {
 338     if (val == (char)val) {
 339         o(0xc483);
 340         g(val);
 341     } else {
 342         oad(0xc481, val); /* add $xxx, %esp */
 343     }
 344 }
 345
 346 /* generate function call with address in (vtop->t, vtop->c) and free function
 347    context. Stack entry is popped */
 348 void gfunc_call(GFuncContext *c)
 349 {
 350     int r;
 351     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 352         /* constant case */
 353         if (vtop->r & VT_SYM) {
 354             /* relocation case */
 355             greloc(cur_text_section, vtop->c.sym,
 356                    ind + 1, R_386_PC32);
 357             oad(0xe8, -4);
 358         } else {
 359             oad(0xe8, vtop->c.ul - ind - 5);
 360         }
 361     } else {
 362         /* otherwise, indirect call */
 363         r = gv(RC_INT);
 364         o(0xff); /* call *r */
 365         o(0xd0 + r);
 366     }
 367     if (c->args_size && c->func_call == FUNC_CDECL)
 368         gadd_sp(c->args_size);
 369     vtop--;
 370 }
 371
 372 /* generate function prolog of type 't' */
 373 void gfunc_prolog(int t)
 374 {
 375     int addr, align, size, u, func_call;
 376     Sym *sym;
 377
 378     sym = sym_find((unsigned)t >> VT_STRUCT_SHIFT);
 379     func_call = sym->r;
 380     addr = 8;
 381     /* if the function returns a structure, then add an
 382        implicit pointer parameter */
 383     func_vt = sym->t;
 384     if ((func_vt & VT_BTYPE) == VT_STRUCT) {
 385         func_vc = addr;
 386         addr += 4;
 387     }
 388     /* define parameters */
 389     while ((sym = sym->next) != NULL) {
 390         u = sym->t;
 391         sym_push(sym->v & ~SYM_FIELD, u,
 392                  VT_LOCAL | VT_LVAL, addr);
 393         size = type_size(u, &align);
 394         size = (size + 3) & ~3;
 395 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 396         /* structs are passed as pointer */
 397         if ((u & VT_BTYPE) == VT_STRUCT) {
 398             size = 4;
 399         }
 400 #endif
 401         addr += size;
 402     }
 403     func_ret_sub = 0;
 404     /* pascal type call ? */
 405     if (func_call == FUNC_STDCALL)
 406         func_ret_sub = addr - 8;
 407     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 408     func_sub_sp_offset = oad(0xec81, 0); /* sub $xxx, %esp */
 409     /* leave some room for bound checking code */
 410     if (do_bounds_check) {
 411         oad(0xb8, 0); /* lbound section pointer */
 412         oad(0xb8, 0); /* call to function */
 413         func_bound_offset = lbounds_section->data_offset;
 414     }
 415 }
 416
 417 /* generate function epilog */
 418 void gfunc_epilog(void)
 419 {
 420 #ifdef CONFIG_TCC_BCHECK
 421     if (do_bounds_check && func_bound_offset != lbounds_section->data_offset) {
 422         int saved_ind;
 423         int *bounds_ptr;
 424         Sym *sym, *sym_data;
 425         /* add end of table info */
 426         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 427         *bounds_ptr = 0;
 428         /* generate bound local allocation */
 429         saved_ind = ind;
 430         ind = func_sub_sp_offset + 4;
 431         sym_data = get_sym_ref(char_pointer_type, lbounds_section,
 432                                func_bound_offset, lbounds_section->data_offset);
 433         greloc(cur_text_section, sym_data,
 434                ind + 1, 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, 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, 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, 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     *(int *)(cur_text_section->data + func_sub_sp_offset) = (-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 *)(cur_text_section->data + *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, 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, 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, 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.ul = (cur_text_section->reloc->data_offset - 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 = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
 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