i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001-2004 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(unsigned int c)
 104 {
 105     while (c) {
 106         g(c);
 107         c = c >> 8;
 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 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 324
 325 /* Generate function call. The function address is pushed first, then
 326    all the parameters in call order. This functions pops all the
 327    parameters and the function address. */
 328 void gfunc_call(int nb_args)
 329 {
 330     int size, align, r, args_size, i, func_call;
 331     Sym *func_sym;
 332
 333     args_size = 0;
 334     for(i = 0;i < nb_args; i++) {
 335         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 336             size = type_size(&vtop->type, &align);
 337             /* align to stack align size */
 338             size = (size + 3) & ~3;
 339             /* allocate the necessary size on stack */
 340             oad(0xec81, size); /* sub $xxx, %esp */
 341             /* generate structure store */
 342             r = get_reg(RC_INT);
 343             o(0x89); /* mov %esp, r */
 344             o(0xe0 + r);
 345             vset(&vtop->type, r | VT_LVAL, 0);
 346             vswap();
 347             vstore();
 348             args_size += size;
 349         } else if (is_float(vtop->type.t)) {
 350             gv(RC_FLOAT); /* only one float register */
 351             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 352                 size = 4;
 353             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 354                 size = 8;
 355             else
 356                 size = 12;
 357             oad(0xec81, size); /* sub $xxx, %esp */
 358             if (size == 12)
 359                 o(0x7cdb);
 360             else
 361                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 362             g(0x24);
 363             g(0x00);
 364             args_size += size;
 365         } else {
 366             /* simple type (currently always same size) */
 367             /* XXX: implicit cast ? */
 368             r = gv(RC_INT);
 369             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 370                 size = 8;
 371                 o(0x50 + vtop->r2); /* push r */
 372             } else {
 373                 size = 4;
 374             }
 375             o(0x50 + r); /* push r */
 376             args_size += size;
 377         }
 378         vtop--;
 379     }
 380     save_regs(0); /* save used temporary registers */
 381     func_sym = vtop->type.ref;
 382     func_call = func_sym->r;
 383     /* fast call case */
 384     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 385         int fastcall_nb_regs;
 386         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 387         for(i = 0;i < fastcall_nb_regs; i++) {
 388             if (args_size <= 0)
 389                 break;
 390             o(0x58 + fastcall_regs[i]); /* pop r */
 391             /* XXX: incorrect for struct/floats */
 392             args_size -= 4;
 393         }
 394     }
 395     gcall_or_jmp(0);
 396     if (args_size && func_sym->r != FUNC_STDCALL)
 397         gadd_sp(args_size);
 398     vtop--;
 399 }
 400
 401 /* generate function prolog of type 't' */
 402 void gfunc_prolog(CType *func_type)
 403 {
 404     int addr, align, size, func_call, fastcall_nb_regs;
 405     int param_index, param_addr;
 406     Sym *sym;
 407     CType *type;
 408
 409     sym = func_type->ref;
 410     func_call = sym->r;
 411     addr = 8;
 412     loc = 0;
 413     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 414         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 415     } else {
 416         fastcall_nb_regs = 0;
 417     }
 418     param_index = 0;
 419
 420     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 421     func_sub_sp_offset = oad(0xec81, 0); /* sub $xxx, %esp */
 422
 423     /* if the function returns a structure, then add an
 424        implicit pointer parameter */
 425     func_vt = sym->type;
 426     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 427         /* XXX: fastcall case ? */
 428         func_vc = addr;
 429         addr += 4;
 430         param_index++;
 431     }
 432     /* define parameters */
 433     while ((sym = sym->next) != NULL) {
 434         type = &sym->type;
 435         size = type_size(type, &align);
 436         size = (size + 3) & ~3;
 437 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 438         /* structs are passed as pointer */
 439         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 440             size = 4;
 441         }
 442 #endif
 443         if (param_index < fastcall_nb_regs) {
 444             /* save FASTCALL register */
 445             loc -= 4;
 446             o(0x89);     /* movl */
 447             gen_modrm(fastcall_regs[param_index], VT_LOCAL, NULL, loc);
 448             param_addr = loc;
 449         } else {
 450             param_addr = addr;
 451             addr += size;
 452         }
 453         sym_push(sym->v & ~SYM_FIELD, type,
 454                  VT_LOCAL | VT_LVAL, param_addr);
 455         param_index++;
 456     }
 457     func_ret_sub = 0;
 458     /* pascal type call ? */
 459     if (func_call == FUNC_STDCALL)
 460         func_ret_sub = addr - 8;
 461
 462     /* leave some room for bound checking code */
 463     if (do_bounds_check) {
 464         oad(0xb8, 0); /* lbound section pointer */
 465         oad(0xb8, 0); /* call to function */
 466         func_bound_offset = lbounds_section->data_offset;
 467     }
 468 }
 469
 470 /* generate function epilog */
 471 void gfunc_epilog(void)
 472 {
 473 #ifdef CONFIG_TCC_BCHECK
 474     if (do_bounds_check && func_bound_offset != lbounds_section->data_offset) {
 475         int saved_ind;
 476         int *bounds_ptr;
 477         Sym *sym, *sym_data;
 478         /* add end of table info */
 479         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 480         *bounds_ptr = 0;
 481         /* generate bound local allocation */
 482         saved_ind = ind;
 483         ind = func_sub_sp_offset + 4;
 484         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 485                                func_bound_offset, lbounds_section->data_offset);
 486         greloc(cur_text_section, sym_data,
 487                ind + 1, R_386_32);
 488         oad(0xb8, 0); /* mov %eax, xxx */
 489         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 490         greloc(cur_text_section, sym,
 491                ind + 1, R_386_PC32);
 492         oad(0xe8, -4);
 493         ind = saved_ind;
 494         /* generate bound check local freeing */
 495         o(0x5250); /* save returned value, if any */
 496         greloc(cur_text_section, sym_data,
 497                ind + 1, R_386_32);
 498         oad(0xb8, 0); /* mov %eax, xxx */
 499         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 500         greloc(cur_text_section, sym,
 501                ind + 1, R_386_PC32);
 502         oad(0xe8, -4);
 503         o(0x585a); /* restore returned value, if any */
 504     }
 505 #endif
 506     o(0xc9); /* leave */
 507     if (func_ret_sub == 0) {
 508         o(0xc3); /* ret */
 509     } else {
 510         o(0xc2); /* ret n */
 511         g(func_ret_sub);
 512         g(func_ret_sub >> 8);
 513     }
 514     /* align local size to word & save local variables */
 515     *(int *)(cur_text_section->data + func_sub_sp_offset) = (-loc + 3) & -4;
 516 }
 517
 518 /* generate a jump to a label */
 519 int gjmp(int t)
 520 {
 521     return psym(0xe9, t);
 522 }
 523
 524 /* generate a jump to a fixed address */
 525 void gjmp_addr(int a)
 526 {
 527     int r;
 528     r = a - ind - 2;
 529     if (r == (char)r) {
 530         g(0xeb);
 531         g(r);
 532     } else {
 533         oad(0xe9, a - ind - 5);
 534     }
 535 }
 536
 537 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 538 int gtst(int inv, int t)
 539 {
 540     int v, *p;
 541
 542     v = vtop->r & VT_VALMASK;
 543     if (v == VT_CMP) {
 544         /* fast case : can jump directly since flags are set */
 545         g(0x0f);
 546         t = psym((vtop->c.i - 16) ^ inv, t);
 547     } else if (v == VT_JMP || v == VT_JMPI) {
 548         /* && or || optimization */
 549         if ((v & 1) == inv) {
 550             /* insert vtop->c jump list in t */
 551             p = &vtop->c.i;
 552             while (*p != 0)
 553                 p = (int *)(cur_text_section->data + *p);
 554             *p = t;
 555             t = vtop->c.i;
 556         } else {
 557             t = gjmp(t);
 558             gsym(vtop->c.i);
 559         }
 560     } else {
 561         if (is_float(vtop->type.t)) {
 562             vpushi(0);
 563             gen_op(TOK_NE);
 564         }
 565         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 566             /* constant jmp optimization */
 567             if ((vtop->c.i != 0) != inv)
 568                 t = gjmp(t);
 569         } else {
 570             v = gv(RC_INT);
 571             o(0x85);
 572             o(0xc0 + v * 9);
 573             g(0x0f);
 574             t = psym(0x85 ^ inv, t);
 575         }
 576     }
 577     vtop--;
 578     return t;
 579 }
 580
 581 /* generate an integer binary operation */
 582 void gen_opi(int op)
 583 {
 584     int r, fr, opc, c;
 585
 586     switch(op) {
 587     case '+':
 588     case TOK_ADDC1: /* add with carry generation */
 589         opc = 0;
 590     gen_op8:
 591         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 592             /* constant case */
 593             vswap();
 594             r = gv(RC_INT);
 595             vswap();
 596             c = vtop->c.i;
 597             if (c == (char)c) {
 598                 /* XXX: generate inc and dec for smaller code ? */
 599                 o(0x83);
 600                 o(0xc0 | (opc << 3) | r);
 601                 g(c);
 602             } else {
 603                 o(0x81);
 604                 oad(0xc0 | (opc << 3) | r, c);
 605             }
 606         } else {
 607             gv2(RC_INT, RC_INT);
 608             r = vtop[-1].r;
 609             fr = vtop[0].r;
 610             o((opc << 3) | 0x01);
 611             o(0xc0 + r + fr * 8);
 612         }
 613         vtop--;
 614         if (op >= TOK_ULT && op <= TOK_GT) {
 615             vtop->r = VT_CMP;
 616             vtop->c.i = op;
 617         }
 618         break;
 619     case '-':
 620     case TOK_SUBC1: /* sub with carry generation */
 621         opc = 5;
 622         goto gen_op8;
 623     case TOK_ADDC2: /* add with carry use */
 624         opc = 2;
 625         goto gen_op8;
 626     case TOK_SUBC2: /* sub with carry use */
 627         opc = 3;
 628         goto gen_op8;
 629     case '&':
 630         opc = 4;
 631         goto gen_op8;
 632     case '^':
 633         opc = 6;
 634         goto gen_op8;
 635     case '|':
 636         opc = 1;
 637         goto gen_op8;
 638     case '*':
 639         gv2(RC_INT, RC_INT);
 640         r = vtop[-1].r;
 641         fr = vtop[0].r;
 642         vtop--;
 643         o(0xaf0f); /* imul fr, r */
 644         o(0xc0 + fr + r * 8);
 645         break;
 646     case TOK_SHL:
 647         opc = 4;
 648         goto gen_shift;
 649     case TOK_SHR:
 650         opc = 5;
 651         goto gen_shift;
 652     case TOK_SAR:
 653         opc = 7;
 654     gen_shift:
 655         opc = 0xc0 | (opc << 3);
 656         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 657             /* constant case */
 658             vswap();
 659             r = gv(RC_INT);
 660             vswap();
 661             c = vtop->c.i & 0x1f;
 662             o(0xc1); /* shl/shr/sar $xxx, r */
 663             o(opc | r);
 664             g(c);
 665         } else {
 666             /* we generate the shift in ecx */
 667             gv2(RC_INT, RC_ECX);
 668             r = vtop[-1].r;
 669             o(0xd3); /* shl/shr/sar %cl, r */
 670             o(opc | r);
 671         }
 672         vtop--;
 673         break;
 674     case '/':
 675     case TOK_UDIV:
 676     case TOK_PDIV:
 677     case '%':
 678     case TOK_UMOD:
 679     case TOK_UMULL:
 680         /* first operand must be in eax */
 681         /* XXX: need better constraint for second operand */
 682         gv2(RC_EAX, RC_ECX);
 683         r = vtop[-1].r;
 684         fr = vtop[0].r;
 685         vtop--;
 686         save_reg(TREG_EDX);
 687         if (op == TOK_UMULL) {
 688             o(0xf7); /* mul fr */
 689             o(0xe0 + fr);
 690             vtop->r2 = TREG_EDX;
 691             r = TREG_EAX;
 692         } else {
 693             if (op == TOK_UDIV || op == TOK_UMOD) {
 694                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 695                 o(0xf0 + fr);
 696             } else {
 697                 o(0xf799); /* cltd, idiv fr, %eax */
 698                 o(0xf8 + fr);
 699             }
 700             if (op == '%' || op == TOK_UMOD)
 701                 r = TREG_EDX;
 702             else
 703                 r = TREG_EAX;
 704         }
 705         vtop->r = r;
 706         break;
 707     default:
 708         opc = 7;
 709         goto gen_op8;
 710     }
 711 }
 712
 713 /* generate a floating point operation 'v = t1 op t2' instruction. The
 714    two operands are guaranted to have the same floating point type */
 715 /* XXX: need to use ST1 too */
 716 void gen_opf(int op)
 717 {
 718     int a, ft, fc, swapped, r;
 719
 720     /* convert constants to memory references */
 721     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 722         vswap();
 723         gv(RC_FLOAT);
 724         vswap();
 725     }
 726     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 727         gv(RC_FLOAT);
 728
 729     /* must put at least one value in the floating point register */
 730     if ((vtop[-1].r & VT_LVAL) &&
 731         (vtop[0].r & VT_LVAL)) {
 732         vswap();
 733         gv(RC_FLOAT);
 734         vswap();
 735     }
 736     swapped = 0;
 737     /* swap the stack if needed so that t1 is the register and t2 is
 738        the memory reference */
 739     if (vtop[-1].r & VT_LVAL) {
 740         vswap();
 741         swapped = 1;
 742     }
 743     if (op >= TOK_ULT && op <= TOK_GT) {
 744         /* load on stack second operand */
 745         load(TREG_ST0, vtop);
 746         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 747         if (op == TOK_GE || op == TOK_GT)
 748             swapped = !swapped;
 749         else if (op == TOK_EQ || op == TOK_NE)
 750             swapped = 0;
 751         if (swapped)
 752             o(0xc9d9); /* fxch %st(1) */
 753         o(0xe9da); /* fucompp */
 754         o(0xe0df); /* fnstsw %ax */
 755         if (op == TOK_EQ) {
 756             o(0x45e480); /* and $0x45, %ah */
 757             o(0x40fC80); /* cmp $0x40, %ah */
 758         } else if (op == TOK_NE) {
 759             o(0x45e480); /* and $0x45, %ah */
 760             o(0x40f480); /* xor $0x40, %ah */
 761             op = TOK_NE;
 762         } else if (op == TOK_GE || op == TOK_LE) {
 763             o(0x05c4f6); /* test $0x05, %ah */
 764             op = TOK_EQ;
 765         } else {
 766             o(0x45c4f6); /* test $0x45, %ah */
 767             op = TOK_EQ;
 768         }
 769         vtop--;
 770         vtop->r = VT_CMP;
 771         vtop->c.i = op;
 772     } else {
 773         /* no memory reference possible for long double operations */
 774         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 775             load(TREG_ST0, vtop);
 776             swapped = !swapped;
 777         }
 778
 779         switch(op) {
 780         default:
 781         case '+':
 782             a = 0;
 783             break;
 784         case '-':
 785             a = 4;
 786             if (swapped)
 787                 a++;
 788             break;
 789         case '*':
 790             a = 1;
 791             break;
 792         case '/':
 793             a = 6;
 794             if (swapped)
 795                 a++;
 796             break;
 797         }
 798         ft = vtop->type.t;
 799         fc = vtop->c.ul;
 800         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 801             o(0xde); /* fxxxp %st, %st(1) */
 802             o(0xc1 + (a << 3));
 803         } else {
 804             /* if saved lvalue, then we must reload it */
 805             r = vtop->r;
 806             if ((r & VT_VALMASK) == VT_LLOCAL) {
 807                 SValue v1;
 808                 r = get_reg(RC_INT);
 809                 v1.type.t = VT_INT;
 810                 v1.r = VT_LOCAL | VT_LVAL;
 811                 v1.c.ul = fc;
 812                 load(r, &v1);
 813                 fc = 0;
 814             }
 815
 816             if ((ft & VT_BTYPE) == VT_DOUBLE)
 817                 o(0xdc);
 818             else
 819                 o(0xd8);
 820             gen_modrm(a, r, vtop->sym, fc);
 821         }
 822         vtop--;
 823     }
 824 }
 825
 826 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 827    and 'long long' cases. */
 828 void gen_cvt_itof(int t)
 829 {
 830     save_reg(TREG_ST0);
 831     gv(RC_INT);
 832     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 833         /* signed long long to float/double/long double (unsigned case
 834            is handled generically) */
 835         o(0x50 + vtop->r2); /* push r2 */
 836         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 837         o(0x242cdf); /* fildll (%esp) */
 838         o(0x08c483); /* add $8, %esp */
 839     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 840                (VT_INT | VT_UNSIGNED)) {
 841         /* unsigned int to float/double/long double */
 842         o(0x6a); /* push $0 */
 843         g(0x00);
 844         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 845         o(0x242cdf); /* fildll (%esp) */
 846         o(0x08c483); /* add $8, %esp */
 847     } else {
 848         /* int to float/double/long double */
 849         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 850         o(0x2404db); /* fildl (%esp) */
 851         o(0x04c483); /* add $4, %esp */
 852     }
 853     vtop->r = TREG_ST0;
 854 }
 855
 856 /* convert fp to int 't' type */
 857 /* XXX: handle long long case */
 858 void gen_cvt_ftoi(int t)
 859 {
 860     int r, r2, size;
 861     Sym *sym;
 862     CType ushort_type;
 863
 864     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 865
 866     gv(RC_FLOAT);
 867     if (t != VT_INT)
 868         size = 8;
 869     else
 870         size = 4;
 871
 872     o(0x2dd9); /* ldcw xxx */
 873     sym = external_global_sym(TOK___tcc_int_fpu_control,
 874                               &ushort_type, VT_LVAL);
 875     greloc(cur_text_section, sym,
 876            ind, R_386_32);
 877     gen_le32(0);
 878
 879     oad(0xec81, size); /* sub $xxx, %esp */
 880     if (size == 4)
 881         o(0x1cdb); /* fistpl */
 882     else
 883         o(0x3cdf); /* fistpll */
 884     o(0x24);
 885     o(0x2dd9); /* ldcw xxx */
 886     sym = external_global_sym(TOK___tcc_fpu_control,
 887                               &ushort_type, VT_LVAL);
 888     greloc(cur_text_section, sym,
 889            ind, R_386_32);
 890     gen_le32(0);
 891
 892     r = get_reg(RC_INT);
 893     o(0x58 + r); /* pop r */
 894     if (size == 8) {
 895         if (t == VT_LLONG) {
 896             vtop->r = r; /* mark reg as used */
 897             r2 = get_reg(RC_INT);
 898             o(0x58 + r2); /* pop r2 */
 899             vtop->r2 = r2;
 900         } else {
 901             o(0x04c483); /* add $4, %esp */
 902         }
 903     }
 904     vtop->r = r;
 905 }
 906
 907 /* convert from one floating point type to another */
 908 void gen_cvt_ftof(int t)
 909 {
 910     /* all we have to do on i386 is to put the float in a register */
 911     gv(RC_FLOAT);
 912 }
 913
 914 /* computed goto support */
 915 void ggoto(void)
 916 {
 917     gcall_or_jmp(1);
 918     vtop--;
 919 }
 920
 921 /* bound check support functions */
 922 #ifdef CONFIG_TCC_BCHECK
 923
 924 /* generate a bounded pointer addition */
 925 void gen_bounded_ptr_add(void)
 926 {
 927     Sym *sym;
 928
 929     /* prepare fast i386 function call (args in eax and edx) */
 930     gv2(RC_EAX, RC_EDX);
 931     /* save all temporary registers */
 932     vtop -= 2;
 933     save_regs(0);
 934     /* do a fast function call */
 935     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
 936     greloc(cur_text_section, sym,
 937            ind + 1, R_386_PC32);
 938     oad(0xe8, -4);
 939     /* returned pointer is in eax */
 940     vtop++;
 941     vtop->r = TREG_EAX | VT_BOUNDED;
 942     /* address of bounding function call point */
 943     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 944 }
 945
 946 /* patch pointer addition in vtop so that pointer dereferencing is
 947    also tested */
 948 void gen_bounded_ptr_deref(void)
 949 {
 950     int func;
 951     int size, align;
 952     Elf32_Rel *rel;
 953     Sym *sym;
 954
 955     size = 0;
 956     /* XXX: put that code in generic part of tcc */
 957     if (!is_float(vtop->type.t)) {
 958         if (vtop->r & VT_LVAL_BYTE)
 959             size = 1;
 960         else if (vtop->r & VT_LVAL_SHORT)
 961             size = 2;
 962     }
 963     if (!size)
 964         size = type_size(&vtop->type, &align);
 965     switch(size) {
 966     case  1: func = TOK___bound_ptr_indir1; break;
 967     case  2: func = TOK___bound_ptr_indir2; break;
 968     case  4: func = TOK___bound_ptr_indir4; break;
 969     case  8: func = TOK___bound_ptr_indir8; break;
 970     case 12: func = TOK___bound_ptr_indir12; break;
 971     case 16: func = TOK___bound_ptr_indir16; break;
 972     default:
 973         error("unhandled size when derefencing bounded pointer");
 974         func = 0;
 975         break;
 976     }
 977
 978     /* patch relocation */
 979     /* XXX: find a better solution ? */
 980     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
 981     sym = external_global_sym(func, &func_old_type, 0);
 982     if (!sym->c)
 983         put_extern_sym(sym, NULL, 0, 0);
 984     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
 985 }
 986 #endif
 987
 988 /* end of X86 code generator */
 989 /*************************************************************/
 990