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 || bt == VT_BOOL)
 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 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 325
 326 /* Generate function call. The function address is pushed first, then
 327    all the parameters in call order. This functions pops all the
 328    parameters and the function address. */
 329 void gfunc_call(int nb_args)
 330 {
 331     int size, align, r, args_size, i, func_call;
 332     Sym *func_sym;
 333
 334     args_size = 0;
 335     for(i = 0;i < nb_args; i++) {
 336         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 337             size = type_size(&vtop->type, &align);
 338             /* align to stack align size */
 339             size = (size + 3) & ~3;
 340             /* allocate the necessary size on stack */
 341             oad(0xec81, size); /* sub $xxx, %esp */
 342             /* generate structure store */
 343             r = get_reg(RC_INT);
 344             o(0x89); /* mov %esp, r */
 345             o(0xe0 + r);
 346             vset(&vtop->type, r | VT_LVAL, 0);
 347             vswap();
 348             vstore();
 349             args_size += size;
 350         } else if (is_float(vtop->type.t)) {
 351             gv(RC_FLOAT); /* only one float register */
 352             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 353                 size = 4;
 354             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 355                 size = 8;
 356             else
 357                 size = 12;
 358             oad(0xec81, size); /* sub $xxx, %esp */
 359             if (size == 12)
 360                 o(0x7cdb);
 361             else
 362                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 363             g(0x24);
 364             g(0x00);
 365             args_size += size;
 366         } else {
 367             /* simple type (currently always same size) */
 368             /* XXX: implicit cast ? */
 369             r = gv(RC_INT);
 370             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 371                 size = 8;
 372                 o(0x50 + vtop->r2); /* push r */
 373             } else {
 374                 size = 4;
 375             }
 376             o(0x50 + r); /* push r */
 377             args_size += size;
 378         }
 379         vtop--;
 380     }
 381     save_regs(0); /* save used temporary registers */
 382     func_sym = vtop->type.ref;
 383     func_call = FUNC_CALL(func_sym->r);
 384     /* fast call case */
 385     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 386         func_call == FUNC_FASTCALLW) {
 387         int fastcall_nb_regs;
 388         uint8_t *fastcall_regs_ptr;
 389         if (func_call == FUNC_FASTCALLW) {
 390             fastcall_regs_ptr = fastcallw_regs;
 391             fastcall_nb_regs = 2;
 392         } else {
 393             fastcall_regs_ptr = fastcall_regs;
 394             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 395         }
 396         for(i = 0;i < fastcall_nb_regs; i++) {
 397             if (args_size <= 0)
 398                 break;
 399             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 400             /* XXX: incorrect for struct/floats */
 401             args_size -= 4;
 402         }
 403     }
 404     gcall_or_jmp(0);
 405     if (args_size && func_call != FUNC_STDCALL)
 406         gadd_sp(args_size);
 407     vtop--;
 408 }
 409
 410 #ifdef TCC_TARGET_PE
 411 #define FUNC_PROLOG_SIZE 10
 412 #else
 413 #define FUNC_PROLOG_SIZE 9
 414 #endif
 415
 416 /* generate function prolog of type 't' */
 417 void gfunc_prolog(CType *func_type)
 418 {
 419     int addr, align, size, func_call, fastcall_nb_regs;
 420     int param_index, param_addr;
 421     uint8_t *fastcall_regs_ptr;
 422     Sym *sym;
 423     CType *type;
 424
 425     sym = func_type->ref;
 426     func_call = FUNC_CALL(sym->r);
 427     addr = 8;
 428     loc = 0;
 429     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 430         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 431         fastcall_regs_ptr = fastcall_regs;
 432     } else if (func_call == FUNC_FASTCALLW) {
 433         fastcall_nb_regs = 2;
 434         fastcall_regs_ptr = fastcallw_regs;
 435     } else {
 436         fastcall_nb_regs = 0;
 437         fastcall_regs_ptr = NULL;
 438     }
 439     param_index = 0;
 440
 441     ind += FUNC_PROLOG_SIZE;
 442     func_sub_sp_offset = ind;
 443     /* if the function returns a structure, then add an
 444        implicit pointer parameter */
 445     func_vt = sym->type;
 446     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 447         /* XXX: fastcall case ? */
 448         func_vc = addr;
 449         addr += 4;
 450         param_index++;
 451     }
 452     /* define parameters */
 453     while ((sym = sym->next) != NULL) {
 454         type = &sym->type;
 455         size = type_size(type, &align);
 456         size = (size + 3) & ~3;
 457 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 458         /* structs are passed as pointer */
 459         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 460             size = 4;
 461         }
 462 #endif
 463         if (param_index < fastcall_nb_regs) {
 464             /* save FASTCALL register */
 465             loc -= 4;
 466             o(0x89);     /* movl */
 467             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 468             param_addr = loc;
 469         } else {
 470             param_addr = addr;
 471             addr += size;
 472         }
 473         sym_push(sym->v & ~SYM_FIELD, type,
 474                  VT_LOCAL | lvalue_type(type->t), param_addr);
 475         param_index++;
 476     }
 477     func_ret_sub = 0;
 478     /* pascal type call ? */
 479     if (func_call == FUNC_STDCALL)
 480         func_ret_sub = addr - 8;
 481
 482     /* leave some room for bound checking code */
 483     if (tcc_state->do_bounds_check) {
 484         oad(0xb8, 0); /* lbound section pointer */
 485         oad(0xb8, 0); /* call to function */
 486         func_bound_offset = lbounds_section->data_offset;
 487     }
 488 }
 489
 490 /* generate function epilog */
 491 void gfunc_epilog(void)
 492 {
 493     int v, saved_ind;
 494
 495 #ifdef CONFIG_TCC_BCHECK
 496     if (tcc_state->do_bounds_check
 497      && func_bound_offset != lbounds_section->data_offset) {
 498         int saved_ind;
 499         int *bounds_ptr;
 500         Sym *sym, *sym_data;
 501         /* add end of table info */
 502         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 503         *bounds_ptr = 0;
 504         /* generate bound local allocation */
 505         saved_ind = ind;
 506         ind = func_sub_sp_offset;
 507         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 508                                func_bound_offset, lbounds_section->data_offset);
 509         greloc(cur_text_section, sym_data,
 510                ind + 1, R_386_32);
 511         oad(0xb8, 0); /* mov %eax, xxx */
 512         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 513         greloc(cur_text_section, sym,
 514                ind + 1, R_386_PC32);
 515         oad(0xe8, -4);
 516         ind = saved_ind;
 517         /* generate bound check local freeing */
 518         o(0x5250); /* save returned value, if any */
 519         greloc(cur_text_section, sym_data,
 520                ind + 1, R_386_32);
 521         oad(0xb8, 0); /* mov %eax, xxx */
 522         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 523         greloc(cur_text_section, sym,
 524                ind + 1, R_386_PC32);
 525         oad(0xe8, -4);
 526         o(0x585a); /* restore returned value, if any */
 527     }
 528 #endif
 529     o(0xc9); /* leave */
 530     if (func_ret_sub == 0) {
 531         o(0xc3); /* ret */
 532     } else {
 533         o(0xc2); /* ret n */
 534         g(func_ret_sub);
 535         g(func_ret_sub >> 8);
 536     }
 537     /* align local size to word & save local variables */
 538
 539     v = (-loc + 3) & -4;
 540     saved_ind = ind;
 541     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 542 #ifdef TCC_TARGET_PE
 543     if (v >= 4096) {
 544         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 545         oad(0xb8, v); /* mov stacksize, %eax */
 546         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 547         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 548     } else
 549 #endif
 550     {
 551         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 552         o(0xec81);  /* sub esp, stacksize */
 553         gen_le32(v);
 554 #if FUNC_PROLOG_SIZE == 10
 555         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 556 #endif
 557     }
 558     ind = saved_ind;
 559 }
 560
 561 /* generate a jump to a label */
 562 int gjmp(int t)
 563 {
 564     return psym(0xe9, t);
 565 }
 566
 567 /* generate a jump to a fixed address */
 568 void gjmp_addr(int a)
 569 {
 570     int r;
 571     r = a - ind - 2;
 572     if (r == (char)r) {
 573         g(0xeb);
 574         g(r);
 575     } else {
 576         oad(0xe9, a - ind - 5);
 577     }
 578 }
 579
 580 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 581 int gtst(int inv, int t)
 582 {
 583     int v, *p;
 584
 585     v = vtop->r & VT_VALMASK;
 586     if (v == VT_CMP) {
 587         /* fast case : can jump directly since flags are set */
 588         g(0x0f);
 589         t = psym((vtop->c.i - 16) ^ inv, t);
 590     } else if (v == VT_JMP || v == VT_JMPI) {
 591         /* && or || optimization */
 592         if ((v & 1) == inv) {
 593             /* insert vtop->c jump list in t */
 594             p = &vtop->c.i;
 595             while (*p != 0)
 596                 p = (int *)(cur_text_section->data + *p);
 597             *p = t;
 598             t = vtop->c.i;
 599         } else {
 600             t = gjmp(t);
 601             gsym(vtop->c.i);
 602         }
 603     } else {
 604         if (is_float(vtop->type.t) ||
 605             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 606             vpushi(0);
 607             gen_op(TOK_NE);
 608         }
 609         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 610             /* constant jmp optimization */
 611             if ((vtop->c.i != 0) != inv)
 612                 t = gjmp(t);
 613         } else {
 614             v = gv(RC_INT);
 615             o(0x85);
 616             o(0xc0 + v * 9);
 617             g(0x0f);
 618             t = psym(0x85 ^ inv, t);
 619         }
 620     }
 621     vtop--;
 622     return t;
 623 }
 624
 625 /* generate an integer binary operation */
 626 void gen_opi(int op)
 627 {
 628     int r, fr, opc, c;
 629
 630     switch(op) {
 631     case '+':
 632     case TOK_ADDC1: /* add with carry generation */
 633         opc = 0;
 634     gen_op8:
 635         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 636             /* constant case */
 637             vswap();
 638             r = gv(RC_INT);
 639             vswap();
 640             c = vtop->c.i;
 641             if (c == (char)c) {
 642                 /* XXX: generate inc and dec for smaller code ? */
 643                 o(0x83);
 644                 o(0xc0 | (opc << 3) | r);
 645                 g(c);
 646             } else {
 647                 o(0x81);
 648                 oad(0xc0 | (opc << 3) | r, c);
 649             }
 650         } else {
 651             gv2(RC_INT, RC_INT);
 652             r = vtop[-1].r;
 653             fr = vtop[0].r;
 654             o((opc << 3) | 0x01);
 655             o(0xc0 + r + fr * 8);
 656         }
 657         vtop--;
 658         if (op >= TOK_ULT && op <= TOK_GT) {
 659             vtop->r = VT_CMP;
 660             vtop->c.i = op;
 661         }
 662         break;
 663     case '-':
 664     case TOK_SUBC1: /* sub with carry generation */
 665         opc = 5;
 666         goto gen_op8;
 667     case TOK_ADDC2: /* add with carry use */
 668         opc = 2;
 669         goto gen_op8;
 670     case TOK_SUBC2: /* sub with carry use */
 671         opc = 3;
 672         goto gen_op8;
 673     case '&':
 674         opc = 4;
 675         goto gen_op8;
 676     case '^':
 677         opc = 6;
 678         goto gen_op8;
 679     case '|':
 680         opc = 1;
 681         goto gen_op8;
 682     case '*':
 683         gv2(RC_INT, RC_INT);
 684         r = vtop[-1].r;
 685         fr = vtop[0].r;
 686         vtop--;
 687         o(0xaf0f); /* imul fr, r */
 688         o(0xc0 + fr + r * 8);
 689         break;
 690     case TOK_SHL:
 691         opc = 4;
 692         goto gen_shift;
 693     case TOK_SHR:
 694         opc = 5;
 695         goto gen_shift;
 696     case TOK_SAR:
 697         opc = 7;
 698     gen_shift:
 699         opc = 0xc0 | (opc << 3);
 700         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 701             /* constant case */
 702             vswap();
 703             r = gv(RC_INT);
 704             vswap();
 705             c = vtop->c.i & 0x1f;
 706             o(0xc1); /* shl/shr/sar $xxx, r */
 707             o(opc | r);
 708             g(c);
 709         } else {
 710             /* we generate the shift in ecx */
 711             gv2(RC_INT, RC_ECX);
 712             r = vtop[-1].r;
 713             o(0xd3); /* shl/shr/sar %cl, r */
 714             o(opc | r);
 715         }
 716         vtop--;
 717         break;
 718     case '/':
 719     case TOK_UDIV:
 720     case TOK_PDIV:
 721     case '%':
 722     case TOK_UMOD:
 723     case TOK_UMULL:
 724         /* first operand must be in eax */
 725         /* XXX: need better constraint for second operand */
 726         gv2(RC_EAX, RC_ECX);
 727         r = vtop[-1].r;
 728         fr = vtop[0].r;
 729         vtop--;
 730         save_reg(TREG_EDX);
 731         if (op == TOK_UMULL) {
 732             o(0xf7); /* mul fr */
 733             o(0xe0 + fr);
 734             vtop->r2 = TREG_EDX;
 735             r = TREG_EAX;
 736         } else {
 737             if (op == TOK_UDIV || op == TOK_UMOD) {
 738                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 739                 o(0xf0 + fr);
 740             } else {
 741                 o(0xf799); /* cltd, idiv fr, %eax */
 742                 o(0xf8 + fr);
 743             }
 744             if (op == '%' || op == TOK_UMOD)
 745                 r = TREG_EDX;
 746             else
 747                 r = TREG_EAX;
 748         }
 749         vtop->r = r;
 750         break;
 751     default:
 752         opc = 7;
 753         goto gen_op8;
 754     }
 755 }
 756
 757 /* generate a floating point operation 'v = t1 op t2' instruction. The
 758    two operands are guaranted to have the same floating point type */
 759 /* XXX: need to use ST1 too */
 760 void gen_opf(int op)
 761 {
 762     int a, ft, fc, swapped, r;
 763
 764     /* convert constants to memory references */
 765     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 766         vswap();
 767         gv(RC_FLOAT);
 768         vswap();
 769     }
 770     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 771         gv(RC_FLOAT);
 772
 773     /* must put at least one value in the floating point register */
 774     if ((vtop[-1].r & VT_LVAL) &&
 775         (vtop[0].r & VT_LVAL)) {
 776         vswap();
 777         gv(RC_FLOAT);
 778         vswap();
 779     }
 780     swapped = 0;
 781     /* swap the stack if needed so that t1 is the register and t2 is
 782        the memory reference */
 783     if (vtop[-1].r & VT_LVAL) {
 784         vswap();
 785         swapped = 1;
 786     }
 787     if (op >= TOK_ULT && op <= TOK_GT) {
 788         /* load on stack second operand */
 789         load(TREG_ST0, vtop);
 790         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 791         if (op == TOK_GE || op == TOK_GT)
 792             swapped = !swapped;
 793         else if (op == TOK_EQ || op == TOK_NE)
 794             swapped = 0;
 795         if (swapped)
 796             o(0xc9d9); /* fxch %st(1) */
 797         o(0xe9da); /* fucompp */
 798         o(0xe0df); /* fnstsw %ax */
 799         if (op == TOK_EQ) {
 800             o(0x45e480); /* and $0x45, %ah */
 801             o(0x40fC80); /* cmp $0x40, %ah */
 802         } else if (op == TOK_NE) {
 803             o(0x45e480); /* and $0x45, %ah */
 804             o(0x40f480); /* xor $0x40, %ah */
 805             op = TOK_NE;
 806         } else if (op == TOK_GE || op == TOK_LE) {
 807             o(0x05c4f6); /* test $0x05, %ah */
 808             op = TOK_EQ;
 809         } else {
 810             o(0x45c4f6); /* test $0x45, %ah */
 811             op = TOK_EQ;
 812         }
 813         vtop--;
 814         vtop->r = VT_CMP;
 815         vtop->c.i = op;
 816     } else {
 817         /* no memory reference possible for long double operations */
 818         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 819             load(TREG_ST0, vtop);
 820             swapped = !swapped;
 821         }
 822
 823         switch(op) {
 824         default:
 825         case '+':
 826             a = 0;
 827             break;
 828         case '-':
 829             a = 4;
 830             if (swapped)
 831                 a++;
 832             break;
 833         case '*':
 834             a = 1;
 835             break;
 836         case '/':
 837             a = 6;
 838             if (swapped)
 839                 a++;
 840             break;
 841         }
 842         ft = vtop->type.t;
 843         fc = vtop->c.ul;
 844         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 845             o(0xde); /* fxxxp %st, %st(1) */
 846             o(0xc1 + (a << 3));
 847         } else {
 848             /* if saved lvalue, then we must reload it */
 849             r = vtop->r;
 850             if ((r & VT_VALMASK) == VT_LLOCAL) {
 851                 SValue v1;
 852                 r = get_reg(RC_INT);
 853                 v1.type.t = VT_INT;
 854                 v1.r = VT_LOCAL | VT_LVAL;
 855                 v1.c.ul = fc;
 856                 load(r, &v1);
 857                 fc = 0;
 858             }
 859
 860             if ((ft & VT_BTYPE) == VT_DOUBLE)
 861                 o(0xdc);
 862             else
 863                 o(0xd8);
 864             gen_modrm(a, r, vtop->sym, fc);
 865         }
 866         vtop--;
 867     }
 868 }
 869
 870 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 871    and 'long long' cases. */
 872 void gen_cvt_itof(int t)
 873 {
 874     save_reg(TREG_ST0);
 875     gv(RC_INT);
 876     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 877         /* signed long long to float/double/long double (unsigned case
 878            is handled generically) */
 879         o(0x50 + vtop->r2); /* push r2 */
 880         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 881         o(0x242cdf); /* fildll (%esp) */
 882         o(0x08c483); /* add $8, %esp */
 883     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 884                (VT_INT | VT_UNSIGNED)) {
 885         /* unsigned int to float/double/long double */
 886         o(0x6a); /* push $0 */
 887         g(0x00);
 888         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 889         o(0x242cdf); /* fildll (%esp) */
 890         o(0x08c483); /* add $8, %esp */
 891     } else {
 892         /* int to float/double/long double */
 893         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 894         o(0x2404db); /* fildl (%esp) */
 895         o(0x04c483); /* add $4, %esp */
 896     }
 897     vtop->r = TREG_ST0;
 898 }
 899
 900 /* convert fp to int 't' type */
 901 /* XXX: handle long long case */
 902 void gen_cvt_ftoi(int t)
 903 {
 904     int r, r2, size;
 905     Sym *sym;
 906     CType ushort_type;
 907
 908     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 909
 910     gv(RC_FLOAT);
 911     if (t != VT_INT)
 912         size = 8;
 913     else
 914         size = 4;
 915
 916     o(0x2dd9); /* ldcw xxx */
 917     sym = external_global_sym(TOK___tcc_int_fpu_control,
 918                               &ushort_type, VT_LVAL);
 919     greloc(cur_text_section, sym,
 920            ind, R_386_32);
 921     gen_le32(0);
 922
 923     oad(0xec81, size); /* sub $xxx, %esp */
 924     if (size == 4)
 925         o(0x1cdb); /* fistpl */
 926     else
 927         o(0x3cdf); /* fistpll */
 928     o(0x24);
 929     o(0x2dd9); /* ldcw xxx */
 930     sym = external_global_sym(TOK___tcc_fpu_control,
 931                               &ushort_type, VT_LVAL);
 932     greloc(cur_text_section, sym,
 933            ind, R_386_32);
 934     gen_le32(0);
 935
 936     r = get_reg(RC_INT);
 937     o(0x58 + r); /* pop r */
 938     if (size == 8) {
 939         if (t == VT_LLONG) {
 940             vtop->r = r; /* mark reg as used */
 941             r2 = get_reg(RC_INT);
 942             o(0x58 + r2); /* pop r2 */
 943             vtop->r2 = r2;
 944         } else {
 945             o(0x04c483); /* add $4, %esp */
 946         }
 947     }
 948     vtop->r = r;
 949 }
 950
 951 /* convert from one floating point type to another */
 952 void gen_cvt_ftof(int t)
 953 {
 954     /* all we have to do on i386 is to put the float in a register */
 955     gv(RC_FLOAT);
 956 }
 957
 958 /* computed goto support */
 959 void ggoto(void)
 960 {
 961     gcall_or_jmp(1);
 962     vtop--;
 963 }
 964
 965 /* bound check support functions */
 966 #ifdef CONFIG_TCC_BCHECK
 967
 968 /* generate a bounded pointer addition */
 969 void gen_bounded_ptr_add(void)
 970 {
 971     Sym *sym;
 972
 973     /* prepare fast i386 function call (args in eax and edx) */
 974     gv2(RC_EAX, RC_EDX);
 975     /* save all temporary registers */
 976     vtop -= 2;
 977     save_regs(0);
 978     /* do a fast function call */
 979     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
 980     greloc(cur_text_section, sym,
 981            ind + 1, R_386_PC32);
 982     oad(0xe8, -4);
 983     /* returned pointer is in eax */
 984     vtop++;
 985     vtop->r = TREG_EAX | VT_BOUNDED;
 986     /* address of bounding function call point */
 987     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 988 }
 989
 990 /* patch pointer addition in vtop so that pointer dereferencing is
 991    also tested */
 992 void gen_bounded_ptr_deref(void)
 993 {
 994     int func;
 995     int size, align;
 996     Elf32_Rel *rel;
 997     Sym *sym;
 998
 999     size = 0;
1000     /* XXX: put that code in generic part of tcc */
1001     if (!is_float(vtop->type.t)) {
1002         if (vtop->r & VT_LVAL_BYTE)
1003             size = 1;
1004         else if (vtop->r & VT_LVAL_SHORT)
1005             size = 2;
1006     }
1007     if (!size)
1008         size = type_size(&vtop->type, &align);
1009     switch(size) {
1010     case  1: func = TOK___bound_ptr_indir1; break;
1011     case  2: func = TOK___bound_ptr_indir2; break;
1012     case  4: func = TOK___bound_ptr_indir4; break;
1013     case  8: func = TOK___bound_ptr_indir8; break;
1014     case 12: func = TOK___bound_ptr_indir12; break;
1015     case 16: func = TOK___bound_ptr_indir16; break;
1016     default:
1017         error("unhandled size when derefencing bounded pointer");
1018         func = 0;
1019         break;
1020     }
1021
1022     /* patch relocation */
1023     /* XXX: find a better solution ? */
1024     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1025     sym = external_global_sym(func, &func_old_type, 0);
1026     if (!sym->c)
1027         put_extern_sym(sym, NULL, 0, 0);
1028     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1029 }
1030 #endif
1031
1032 /* end of X86 code generator */
1033 /*************************************************************/
1034