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 const 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_DATA_PTR  R_386_32
  81 #define R_JMP_SLOT  R_386_JMP_SLOT
  82 #define R_COPY      R_386_COPY
  83
  84 #define ELF_START_ADDR 0x08048000
  85 #define ELF_PAGE_SIZE  0x1000
  86
  87 /******************************************************/
  88
  89 static unsigned long func_sub_sp_offset;
  90 static unsigned long func_bound_offset;
  91 static int func_ret_sub;
  92
  93 /* XXX: make it faster ? */
  94 void g(int c)
  95 {
  96     int ind1;
  97     ind1 = ind + 1;
  98     if (ind1 > cur_text_section->data_allocated)
  99         section_realloc(cur_text_section, ind1);
 100     cur_text_section->data[ind] = c;
 101     ind = ind1;
 102 }
 103
 104 void o(unsigned int c)
 105 {
 106     while (c) {
 107         g(c);
 108         c = c >> 8;
 109     }
 110 }
 111
 112 void gen_le32(int c)
 113 {
 114     g(c);
 115     g(c >> 8);
 116     g(c >> 16);
 117     g(c >> 24);
 118 }
 119
 120 /* output a symbol and patch all calls to it */
 121 void gsym_addr(int t, int a)
 122 {
 123     int n, *ptr;
 124     while (t) {
 125         ptr = (int *)(cur_text_section->data + t);
 126         n = *ptr; /* next value */
 127         *ptr = a - t - 4;
 128         t = n;
 129     }
 130 }
 131
 132 void gsym(int t)
 133 {
 134     gsym_addr(t, ind);
 135 }
 136
 137 /* psym is used to put an instruction with a data field which is a
 138    reference to a symbol. It is in fact the same as oad ! */
 139 #define psym oad
 140
 141 /* instruction + 4 bytes data. Return the address of the data */
 142 static int oad(int c, int s)
 143 {
 144     int ind1;
 145
 146     o(c);
 147     ind1 = ind + 4;
 148     if (ind1 > cur_text_section->data_allocated)
 149         section_realloc(cur_text_section, ind1);
 150     *(int *)(cur_text_section->data + ind) = s;
 151     s = ind;
 152     ind = ind1;
 153     return s;
 154 }
 155
 156 /* output constant with relocation if 'r & VT_SYM' is true */
 157 static void gen_addr32(int r, Sym *sym, int c)
 158 {
 159     if (r & VT_SYM)
 160         greloc(cur_text_section, sym, ind, R_386_32);
 161     gen_le32(c);
 162 }
 163
 164 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 165    opcode bits */
 166 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 167 {
 168     op_reg = op_reg << 3;
 169     if ((r & VT_VALMASK) == VT_CONST) {
 170         /* constant memory reference */
 171         o(0x05 | op_reg);
 172         gen_addr32(r, sym, c);
 173     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 174         /* currently, we use only ebp as base */
 175         if (c == (char)c) {
 176             /* short reference */
 177             o(0x45 | op_reg);
 178             g(c);
 179         } else {
 180             oad(0x85 | op_reg, c);
 181         }
 182     } else {
 183         g(0x00 | op_reg | (r & VT_VALMASK));
 184     }
 185 }
 186
 187
 188 /* load 'r' from value 'sv' */
 189 void load(int r, SValue *sv)
 190 {
 191     int v, t, ft, fc, fr;
 192     SValue v1;
 193
 194     fr = sv->r;
 195     ft = sv->type.t;
 196     fc = sv->c.ul;
 197
 198     v = fr & VT_VALMASK;
 199     if (fr & VT_LVAL) {
 200         if (v == VT_LLOCAL) {
 201             v1.type.t = VT_INT;
 202             v1.r = VT_LOCAL | VT_LVAL;
 203             v1.c.ul = fc;
 204             load(r, &v1);
 205             fr = r;
 206         }
 207         if ((ft & VT_BTYPE) == VT_FLOAT) {
 208             o(0xd9); /* flds */
 209             r = 0;
 210         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 211             o(0xdd); /* fldl */
 212             r = 0;
 213         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 214             o(0xdb); /* fldt */
 215             r = 5;
 216         } else if ((ft & VT_TYPE) == VT_BYTE) {
 217             o(0xbe0f);   /* movsbl */
 218         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 219             o(0xb60f);   /* movzbl */
 220         } else if ((ft & VT_TYPE) == VT_SHORT) {
 221             o(0xbf0f);   /* movswl */
 222         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 223             o(0xb70f);   /* movzwl */
 224         } else {
 225             o(0x8b);     /* movl */
 226         }
 227         gen_modrm(r, fr, sv->sym, fc);
 228     } else {
 229         if (v == VT_CONST) {
 230             o(0xb8 + r); /* mov $xx, r */
 231             gen_addr32(fr, sv->sym, fc);
 232         } else if (v == VT_LOCAL) {
 233             o(0x8d); /* lea xxx(%ebp), r */
 234             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 235         } else if (v == VT_CMP) {
 236             oad(0xb8 + r, 0); /* mov $0, r */
 237             o(0x0f); /* setxx %br */
 238             o(fc);
 239             o(0xc0 + r);
 240         } else if (v == VT_JMP || v == VT_JMPI) {
 241             t = v & 1;
 242             oad(0xb8 + r, t); /* mov $1, r */
 243             o(0x05eb); /* jmp after */
 244             gsym(fc);
 245             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 246         } else if (v != r) {
 247             o(0x89);
 248             o(0xc0 + r + v * 8); /* mov v, r */
 249         }
 250     }
 251 }
 252
 253 /* store register 'r' in lvalue 'v' */
 254 void store(int r, SValue *v)
 255 {
 256     int fr, bt, ft, fc;
 257
 258     ft = v->type.t;
 259     fc = v->c.ul;
 260     fr = v->r & VT_VALMASK;
 261     bt = ft & VT_BTYPE;
 262     /* XXX: incorrect if float reg to reg */
 263     if (bt == VT_FLOAT) {
 264         o(0xd9); /* fsts */
 265         r = 2;
 266     } else if (bt == VT_DOUBLE) {
 267         o(0xdd); /* fstpl */
 268         r = 2;
 269     } else if (bt == VT_LDOUBLE) {
 270         o(0xc0d9); /* fld %st(0) */
 271         o(0xdb); /* fstpt */
 272         r = 7;
 273     } else {
 274         if (bt == VT_SHORT)
 275             o(0x66);
 276         if (bt == VT_BYTE || bt == VT_BOOL)
 277             o(0x88);
 278         else
 279             o(0x89);
 280     }
 281     if (fr == VT_CONST ||
 282         fr == VT_LOCAL ||
 283         (v->r & VT_LVAL)) {
 284         gen_modrm(r, v->r, v->sym, fc);
 285     } else if (fr != r) {
 286         o(0xc0 + fr + r * 8); /* mov r, fr */
 287     }
 288 }
 289
 290 static void gadd_sp(int val)
 291 {
 292     if (val == (char)val) {
 293         o(0xc483);
 294         g(val);
 295     } else {
 296         oad(0xc481, val); /* add $xxx, %esp */
 297     }
 298 }
 299
 300 /* 'is_jmp' is '1' if it is a jump */
 301 static void gcall_or_jmp(int is_jmp)
 302 {
 303     int r;
 304     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 305         /* constant case */
 306         if (vtop->r & VT_SYM) {
 307             /* relocation case */
 308             greloc(cur_text_section, vtop->sym,
 309                    ind + 1, R_386_PC32);
 310         } else {
 311             /* put an empty PC32 relocation */
 312             put_elf_reloc(symtab_section, cur_text_section,
 313                           ind + 1, R_386_PC32, 0);
 314         }
 315         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 316     } else {
 317         /* otherwise, indirect call */
 318         r = gv(RC_INT);
 319         o(0xff); /* call/jmp *r */
 320         o(0xd0 + r + (is_jmp << 4));
 321     }
 322 }
 323
 324 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 325 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 326
 327 /* Generate function call. The function address is pushed first, then
 328    all the parameters in call order. This functions pops all the
 329    parameters and the function address. */
 330 void gfunc_call(int nb_args)
 331 {
 332     int size, align, r, args_size, i, func_call;
 333     Sym *func_sym;
 334
 335     args_size = 0;
 336     for(i = 0;i < nb_args; i++) {
 337         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 338             size = type_size(&vtop->type, &align);
 339             /* align to stack align size */
 340             size = (size + 3) & ~3;
 341             /* allocate the necessary size on stack */
 342             oad(0xec81, size); /* sub $xxx, %esp */
 343             /* generate structure store */
 344             r = get_reg(RC_INT);
 345             o(0x89); /* mov %esp, r */
 346             o(0xe0 + r);
 347             vset(&vtop->type, r | VT_LVAL, 0);
 348             vswap();
 349             vstore();
 350             args_size += size;
 351         } else if (is_float(vtop->type.t)) {
 352             gv(RC_FLOAT); /* only one float register */
 353             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 354                 size = 4;
 355             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 356                 size = 8;
 357             else
 358                 size = 12;
 359             oad(0xec81, size); /* sub $xxx, %esp */
 360             if (size == 12)
 361                 o(0x7cdb);
 362             else
 363                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 364             g(0x24);
 365             g(0x00);
 366             args_size += size;
 367         } else {
 368             /* simple type (currently always same size) */
 369             /* XXX: implicit cast ? */
 370             r = gv(RC_INT);
 371             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 372                 size = 8;
 373                 o(0x50 + vtop->r2); /* push r */
 374             } else {
 375                 size = 4;
 376             }
 377             o(0x50 + r); /* push r */
 378             args_size += size;
 379         }
 380         vtop--;
 381     }
 382     save_regs(0); /* save used temporary registers */
 383     func_sym = vtop->type.ref;
 384     func_call = FUNC_CALL(func_sym->r);
 385     /* fast call case */
 386     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 387         func_call == FUNC_FASTCALLW) {
 388         int fastcall_nb_regs;
 389         uint8_t *fastcall_regs_ptr;
 390         if (func_call == FUNC_FASTCALLW) {
 391             fastcall_regs_ptr = fastcallw_regs;
 392             fastcall_nb_regs = 2;
 393         } else {
 394             fastcall_regs_ptr = fastcall_regs;
 395             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 396         }
 397         for(i = 0;i < fastcall_nb_regs; i++) {
 398             if (args_size <= 0)
 399                 break;
 400             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 401             /* XXX: incorrect for struct/floats */
 402             args_size -= 4;
 403         }
 404     }
 405     gcall_or_jmp(0);
 406
 407 #ifdef TCC_TARGET_PE
 408     if ((func_sym->type.t & VT_BTYPE) == VT_STRUCT)
 409         args_size -= 4;
 410 #endif
 411     if (args_size && func_call != FUNC_STDCALL)
 412         gadd_sp(args_size);
 413     vtop--;
 414 }
 415
 416 #ifdef TCC_TARGET_PE
 417 #define FUNC_PROLOG_SIZE 10
 418 #else
 419 #define FUNC_PROLOG_SIZE 9
 420 #endif
 421
 422 /* generate function prolog of type 't' */
 423 void gfunc_prolog(CType *func_type)
 424 {
 425     int addr, align, size, func_call, fastcall_nb_regs;
 426     int param_index, param_addr;
 427     uint8_t *fastcall_regs_ptr;
 428     Sym *sym;
 429     CType *type;
 430
 431     sym = func_type->ref;
 432     func_call = FUNC_CALL(sym->r);
 433     addr = 8;
 434     loc = 0;
 435     func_vc = 0;
 436
 437     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 438         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 439         fastcall_regs_ptr = fastcall_regs;
 440     } else if (func_call == FUNC_FASTCALLW) {
 441         fastcall_nb_regs = 2;
 442         fastcall_regs_ptr = fastcallw_regs;
 443     } else {
 444         fastcall_nb_regs = 0;
 445         fastcall_regs_ptr = NULL;
 446     }
 447     param_index = 0;
 448
 449     ind += FUNC_PROLOG_SIZE;
 450     func_sub_sp_offset = ind;
 451     /* if the function returns a structure, then add an
 452        implicit pointer parameter */
 453     func_vt = sym->type;
 454     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 455         /* XXX: fastcall case ? */
 456         func_vc = addr;
 457         addr += 4;
 458         param_index++;
 459     }
 460     /* define parameters */
 461     while ((sym = sym->next) != NULL) {
 462         type = &sym->type;
 463         size = type_size(type, &align);
 464         size = (size + 3) & ~3;
 465 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 466         /* structs are passed as pointer */
 467         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 468             size = 4;
 469         }
 470 #endif
 471         if (param_index < fastcall_nb_regs) {
 472             /* save FASTCALL register */
 473             loc -= 4;
 474             o(0x89);     /* movl */
 475             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 476             param_addr = loc;
 477         } else {
 478             param_addr = addr;
 479             addr += size;
 480         }
 481         sym_push(sym->v & ~SYM_FIELD, type,
 482                  VT_LOCAL | lvalue_type(type->t), param_addr);
 483         param_index++;
 484     }
 485     func_ret_sub = 0;
 486     /* pascal type call ? */
 487     if (func_call == FUNC_STDCALL)
 488         func_ret_sub = addr - 8;
 489 #ifdef TCC_TARGET_PE
 490     else if (func_vc)
 491         func_ret_sub = 4;
 492 #endif
 493
 494     /* leave some room for bound checking code */
 495     if (tcc_state->do_bounds_check) {
 496         oad(0xb8, 0); /* lbound section pointer */
 497         oad(0xb8, 0); /* call to function */
 498         func_bound_offset = lbounds_section->data_offset;
 499     }
 500 }
 501
 502 /* generate function epilog */
 503 void gfunc_epilog(void)
 504 {
 505     int v, saved_ind;
 506
 507 #ifdef CONFIG_TCC_BCHECK
 508     if (tcc_state->do_bounds_check
 509      && func_bound_offset != lbounds_section->data_offset) {
 510         int saved_ind;
 511         int *bounds_ptr;
 512         Sym *sym, *sym_data;
 513         /* add end of table info */
 514         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 515         *bounds_ptr = 0;
 516         /* generate bound local allocation */
 517         saved_ind = ind;
 518         ind = func_sub_sp_offset;
 519         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 520                                func_bound_offset, lbounds_section->data_offset);
 521         greloc(cur_text_section, sym_data,
 522                ind + 1, R_386_32);
 523         oad(0xb8, 0); /* mov %eax, xxx */
 524         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 525         greloc(cur_text_section, sym,
 526                ind + 1, R_386_PC32);
 527         oad(0xe8, -4);
 528         ind = saved_ind;
 529         /* generate bound check local freeing */
 530         o(0x5250); /* save returned value, if any */
 531         greloc(cur_text_section, sym_data,
 532                ind + 1, R_386_32);
 533         oad(0xb8, 0); /* mov %eax, xxx */
 534         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 535         greloc(cur_text_section, sym,
 536                ind + 1, R_386_PC32);
 537         oad(0xe8, -4);
 538         o(0x585a); /* restore returned value, if any */
 539     }
 540 #endif
 541     o(0xc9); /* leave */
 542     if (func_ret_sub == 0) {
 543         o(0xc3); /* ret */
 544     } else {
 545         o(0xc2); /* ret n */
 546         g(func_ret_sub);
 547         g(func_ret_sub >> 8);
 548     }
 549     /* align local size to word & save local variables */
 550
 551     v = (-loc + 3) & -4;
 552     saved_ind = ind;
 553     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 554 #ifdef TCC_TARGET_PE
 555     if (v >= 4096) {
 556         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 557         oad(0xb8, v); /* mov stacksize, %eax */
 558         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 559         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 560     } else
 561 #endif
 562     {
 563         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 564         o(0xec81);  /* sub esp, stacksize */
 565         gen_le32(v);
 566 #if FUNC_PROLOG_SIZE == 10
 567         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 568 #endif
 569     }
 570     ind = saved_ind;
 571 }
 572
 573 /* generate a jump to a label */
 574 int gjmp(int t)
 575 {
 576     return psym(0xe9, t);
 577 }
 578
 579 /* generate a jump to a fixed address */
 580 void gjmp_addr(int a)
 581 {
 582     int r;
 583     r = a - ind - 2;
 584     if (r == (char)r) {
 585         g(0xeb);
 586         g(r);
 587     } else {
 588         oad(0xe9, a - ind - 5);
 589     }
 590 }
 591
 592 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 593 int gtst(int inv, int t)
 594 {
 595     int v, *p;
 596
 597     v = vtop->r & VT_VALMASK;
 598     if (v == VT_CMP) {
 599         /* fast case : can jump directly since flags are set */
 600         g(0x0f);
 601         t = psym((vtop->c.i - 16) ^ inv, t);
 602     } else if (v == VT_JMP || v == VT_JMPI) {
 603         /* && or || optimization */
 604         if ((v & 1) == inv) {
 605             /* insert vtop->c jump list in t */
 606             p = &vtop->c.i;
 607             while (*p != 0)
 608                 p = (int *)(cur_text_section->data + *p);
 609             *p = t;
 610             t = vtop->c.i;
 611         } else {
 612             t = gjmp(t);
 613             gsym(vtop->c.i);
 614         }
 615     } else {
 616         if (is_float(vtop->type.t) ||
 617             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 618             vpushi(0);
 619             gen_op(TOK_NE);
 620         }
 621         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 622             /* constant jmp optimization */
 623             if ((vtop->c.i != 0) != inv)
 624                 t = gjmp(t);
 625         } else {
 626             v = gv(RC_INT);
 627             o(0x85);
 628             o(0xc0 + v * 9);
 629             g(0x0f);
 630             t = psym(0x85 ^ inv, t);
 631         }
 632     }
 633     vtop--;
 634     return t;
 635 }
 636
 637 /* generate an integer binary operation */
 638 void gen_opi(int op)
 639 {
 640     int r, fr, opc, c;
 641
 642     switch(op) {
 643     case '+':
 644     case TOK_ADDC1: /* add with carry generation */
 645         opc = 0;
 646     gen_op8:
 647         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 648             /* constant case */
 649             vswap();
 650             r = gv(RC_INT);
 651             vswap();
 652             c = vtop->c.i;
 653             if (c == (char)c) {
 654                 /* XXX: generate inc and dec for smaller code ? */
 655                 o(0x83);
 656                 o(0xc0 | (opc << 3) | r);
 657                 g(c);
 658             } else {
 659                 o(0x81);
 660                 oad(0xc0 | (opc << 3) | r, c);
 661             }
 662         } else {
 663             gv2(RC_INT, RC_INT);
 664             r = vtop[-1].r;
 665             fr = vtop[0].r;
 666             o((opc << 3) | 0x01);
 667             o(0xc0 + r + fr * 8);
 668         }
 669         vtop--;
 670         if (op >= TOK_ULT && op <= TOK_GT) {
 671             vtop->r = VT_CMP;
 672             vtop->c.i = op;
 673         }
 674         break;
 675     case '-':
 676     case TOK_SUBC1: /* sub with carry generation */
 677         opc = 5;
 678         goto gen_op8;
 679     case TOK_ADDC2: /* add with carry use */
 680         opc = 2;
 681         goto gen_op8;
 682     case TOK_SUBC2: /* sub with carry use */
 683         opc = 3;
 684         goto gen_op8;
 685     case '&':
 686         opc = 4;
 687         goto gen_op8;
 688     case '^':
 689         opc = 6;
 690         goto gen_op8;
 691     case '|':
 692         opc = 1;
 693         goto gen_op8;
 694     case '*':
 695         gv2(RC_INT, RC_INT);
 696         r = vtop[-1].r;
 697         fr = vtop[0].r;
 698         vtop--;
 699         o(0xaf0f); /* imul fr, r */
 700         o(0xc0 + fr + r * 8);
 701         break;
 702     case TOK_SHL:
 703         opc = 4;
 704         goto gen_shift;
 705     case TOK_SHR:
 706         opc = 5;
 707         goto gen_shift;
 708     case TOK_SAR:
 709         opc = 7;
 710     gen_shift:
 711         opc = 0xc0 | (opc << 3);
 712         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 713             /* constant case */
 714             vswap();
 715             r = gv(RC_INT);
 716             vswap();
 717             c = vtop->c.i & 0x1f;
 718             o(0xc1); /* shl/shr/sar $xxx, r */
 719             o(opc | r);
 720             g(c);
 721         } else {
 722             /* we generate the shift in ecx */
 723             gv2(RC_INT, RC_ECX);
 724             r = vtop[-1].r;
 725             o(0xd3); /* shl/shr/sar %cl, r */
 726             o(opc | r);
 727         }
 728         vtop--;
 729         break;
 730     case '/':
 731     case TOK_UDIV:
 732     case TOK_PDIV:
 733     case '%':
 734     case TOK_UMOD:
 735     case TOK_UMULL:
 736         /* first operand must be in eax */
 737         /* XXX: need better constraint for second operand */
 738         gv2(RC_EAX, RC_ECX);
 739         r = vtop[-1].r;
 740         fr = vtop[0].r;
 741         vtop--;
 742         save_reg(TREG_EDX);
 743         if (op == TOK_UMULL) {
 744             o(0xf7); /* mul fr */
 745             o(0xe0 + fr);
 746             vtop->r2 = TREG_EDX;
 747             r = TREG_EAX;
 748         } else {
 749             if (op == TOK_UDIV || op == TOK_UMOD) {
 750                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 751                 o(0xf0 + fr);
 752             } else {
 753                 o(0xf799); /* cltd, idiv fr, %eax */
 754                 o(0xf8 + fr);
 755             }
 756             if (op == '%' || op == TOK_UMOD)
 757                 r = TREG_EDX;
 758             else
 759                 r = TREG_EAX;
 760         }
 761         vtop->r = r;
 762         break;
 763     default:
 764         opc = 7;
 765         goto gen_op8;
 766     }
 767 }
 768
 769 /* generate a floating point operation 'v = t1 op t2' instruction. The
 770    two operands are guaranted to have the same floating point type */
 771 /* XXX: need to use ST1 too */
 772 void gen_opf(int op)
 773 {
 774     int a, ft, fc, swapped, r;
 775
 776     /* convert constants to memory references */
 777     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 778         vswap();
 779         gv(RC_FLOAT);
 780         vswap();
 781     }
 782     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 783         gv(RC_FLOAT);
 784
 785     /* must put at least one value in the floating point register */
 786     if ((vtop[-1].r & VT_LVAL) &&
 787         (vtop[0].r & VT_LVAL)) {
 788         vswap();
 789         gv(RC_FLOAT);
 790         vswap();
 791     }
 792     swapped = 0;
 793     /* swap the stack if needed so that t1 is the register and t2 is
 794        the memory reference */
 795     if (vtop[-1].r & VT_LVAL) {
 796         vswap();
 797         swapped = 1;
 798     }
 799     if (op >= TOK_ULT && op <= TOK_GT) {
 800         /* load on stack second operand */
 801         load(TREG_ST0, vtop);
 802         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 803         if (op == TOK_GE || op == TOK_GT)
 804             swapped = !swapped;
 805         else if (op == TOK_EQ || op == TOK_NE)
 806             swapped = 0;
 807         if (swapped)
 808             o(0xc9d9); /* fxch %st(1) */
 809         o(0xe9da); /* fucompp */
 810         o(0xe0df); /* fnstsw %ax */
 811         if (op == TOK_EQ) {
 812             o(0x45e480); /* and $0x45, %ah */
 813             o(0x40fC80); /* cmp $0x40, %ah */
 814         } else if (op == TOK_NE) {
 815             o(0x45e480); /* and $0x45, %ah */
 816             o(0x40f480); /* xor $0x40, %ah */
 817             op = TOK_NE;
 818         } else if (op == TOK_GE || op == TOK_LE) {
 819             o(0x05c4f6); /* test $0x05, %ah */
 820             op = TOK_EQ;
 821         } else {
 822             o(0x45c4f6); /* test $0x45, %ah */
 823             op = TOK_EQ;
 824         }
 825         vtop--;
 826         vtop->r = VT_CMP;
 827         vtop->c.i = op;
 828     } else {
 829         /* no memory reference possible for long double operations */
 830         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 831             load(TREG_ST0, vtop);
 832             swapped = !swapped;
 833         }
 834
 835         switch(op) {
 836         default:
 837         case '+':
 838             a = 0;
 839             break;
 840         case '-':
 841             a = 4;
 842             if (swapped)
 843                 a++;
 844             break;
 845         case '*':
 846             a = 1;
 847             break;
 848         case '/':
 849             a = 6;
 850             if (swapped)
 851                 a++;
 852             break;
 853         }
 854         ft = vtop->type.t;
 855         fc = vtop->c.ul;
 856         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 857             o(0xde); /* fxxxp %st, %st(1) */
 858             o(0xc1 + (a << 3));
 859         } else {
 860             /* if saved lvalue, then we must reload it */
 861             r = vtop->r;
 862             if ((r & VT_VALMASK) == VT_LLOCAL) {
 863                 SValue v1;
 864                 r = get_reg(RC_INT);
 865                 v1.type.t = VT_INT;
 866                 v1.r = VT_LOCAL | VT_LVAL;
 867                 v1.c.ul = fc;
 868                 load(r, &v1);
 869                 fc = 0;
 870             }
 871
 872             if ((ft & VT_BTYPE) == VT_DOUBLE)
 873                 o(0xdc);
 874             else
 875                 o(0xd8);
 876             gen_modrm(a, r, vtop->sym, fc);
 877         }
 878         vtop--;
 879     }
 880 }
 881
 882 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 883    and 'long long' cases. */
 884 void gen_cvt_itof(int t)
 885 {
 886     save_reg(TREG_ST0);
 887     gv(RC_INT);
 888     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 889         /* signed long long to float/double/long double (unsigned case
 890            is handled generically) */
 891         o(0x50 + vtop->r2); /* push r2 */
 892         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 893         o(0x242cdf); /* fildll (%esp) */
 894         o(0x08c483); /* add $8, %esp */
 895     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 896                (VT_INT | VT_UNSIGNED)) {
 897         /* unsigned int to float/double/long double */
 898         o(0x6a); /* push $0 */
 899         g(0x00);
 900         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 901         o(0x242cdf); /* fildll (%esp) */
 902         o(0x08c483); /* add $8, %esp */
 903     } else {
 904         /* int to float/double/long double */
 905         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 906         o(0x2404db); /* fildl (%esp) */
 907         o(0x04c483); /* add $4, %esp */
 908     }
 909     vtop->r = TREG_ST0;
 910 }
 911
 912 /* convert fp to int 't' type */
 913 /* XXX: handle long long case */
 914 void gen_cvt_ftoi(int t)
 915 {
 916     int r, r2, size;
 917     Sym *sym;
 918     CType ushort_type;
 919
 920     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 921
 922     gv(RC_FLOAT);
 923     if (t != VT_INT)
 924         size = 8;
 925     else
 926         size = 4;
 927
 928     o(0x2dd9); /* ldcw xxx */
 929     sym = external_global_sym(TOK___tcc_int_fpu_control,
 930                               &ushort_type, VT_LVAL);
 931     greloc(cur_text_section, sym,
 932            ind, R_386_32);
 933     gen_le32(0);
 934
 935     oad(0xec81, size); /* sub $xxx, %esp */
 936     if (size == 4)
 937         o(0x1cdb); /* fistpl */
 938     else
 939         o(0x3cdf); /* fistpll */
 940     o(0x24);
 941     o(0x2dd9); /* ldcw xxx */
 942     sym = external_global_sym(TOK___tcc_fpu_control,
 943                               &ushort_type, VT_LVAL);
 944     greloc(cur_text_section, sym,
 945            ind, R_386_32);
 946     gen_le32(0);
 947
 948     r = get_reg(RC_INT);
 949     o(0x58 + r); /* pop r */
 950     if (size == 8) {
 951         if (t == VT_LLONG) {
 952             vtop->r = r; /* mark reg as used */
 953             r2 = get_reg(RC_INT);
 954             o(0x58 + r2); /* pop r2 */
 955             vtop->r2 = r2;
 956         } else {
 957             o(0x04c483); /* add $4, %esp */
 958         }
 959     }
 960     vtop->r = r;
 961 }
 962
 963 /* convert from one floating point type to another */
 964 void gen_cvt_ftof(int t)
 965 {
 966     /* all we have to do on i386 is to put the float in a register */
 967     gv(RC_FLOAT);
 968 }
 969
 970 /* computed goto support */
 971 void ggoto(void)
 972 {
 973     gcall_or_jmp(1);
 974     vtop--;
 975 }
 976
 977 /* bound check support functions */
 978 #ifdef CONFIG_TCC_BCHECK
 979
 980 /* generate a bounded pointer addition */
 981 void gen_bounded_ptr_add(void)
 982 {
 983     Sym *sym;
 984
 985     /* prepare fast i386 function call (args in eax and edx) */
 986     gv2(RC_EAX, RC_EDX);
 987     /* save all temporary registers */
 988     vtop -= 2;
 989     save_regs(0);
 990     /* do a fast function call */
 991     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
 992     greloc(cur_text_section, sym,
 993            ind + 1, R_386_PC32);
 994     oad(0xe8, -4);
 995     /* returned pointer is in eax */
 996     vtop++;
 997     vtop->r = TREG_EAX | VT_BOUNDED;
 998     /* address of bounding function call point */
 999     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1000 }
1001
1002 /* patch pointer addition in vtop so that pointer dereferencing is
1003    also tested */
1004 void gen_bounded_ptr_deref(void)
1005 {
1006     int func;
1007     int size, align;
1008     Elf32_Rel *rel;
1009     Sym *sym;
1010
1011     size = 0;
1012     /* XXX: put that code in generic part of tcc */
1013     if (!is_float(vtop->type.t)) {
1014         if (vtop->r & VT_LVAL_BYTE)
1015             size = 1;
1016         else if (vtop->r & VT_LVAL_SHORT)
1017             size = 2;
1018     }
1019     if (!size)
1020         size = type_size(&vtop->type, &align);
1021     switch(size) {
1022     case  1: func = TOK___bound_ptr_indir1; break;
1023     case  2: func = TOK___bound_ptr_indir2; break;
1024     case  4: func = TOK___bound_ptr_indir4; break;
1025     case  8: func = TOK___bound_ptr_indir8; break;
1026     case 12: func = TOK___bound_ptr_indir12; break;
1027     case 16: func = TOK___bound_ptr_indir16; break;
1028     default:
1029         error("unhandled size when derefencing bounded pointer");
1030         func = 0;
1031         break;
1032     }
1033
1034     /* patch relocation */
1035     /* XXX: find a better solution ? */
1036     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1037     sym = external_global_sym(func, &func_old_type, 0);
1038     if (!sym->c)
1039         put_extern_sym(sym, NULL, 0, 0);
1040     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1041 }
1042 #endif
1043
1044 /* end of X86 code generator */
1045 /*************************************************************/
1046