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 #ifdef TARGET_DEFS_ONLY
  22
  23 /* number of available registers */
  24 #define NB_REGS         4
  25 #define NB_ASM_REGS     8
  26
  27 /* a register can belong to several classes. The classes must be
  28    sorted from more general to more precise (see gv2() code which does
  29    assumptions on it). */
  30 #define RC_INT     0x0001 /* generic integer register */
  31 #define RC_FLOAT   0x0002 /* generic float register */
  32 #define RC_EAX     0x0004
  33 #define RC_ST0     0x0008
  34 #define RC_ECX     0x0010
  35 #define RC_EDX     0x0020
  36 #define RC_IRET    RC_EAX /* function return: integer register */
  37 #define RC_LRET    RC_EDX /* function return: second integer register */
  38 #define RC_FRET    RC_ST0 /* function return: float register */
  39
  40 /* pretty names for the registers */
  41 enum {
  42     TREG_EAX = 0,
  43     TREG_ECX,
  44     TREG_EDX,
  45     TREG_ST0,
  46 };
  47
  48 /* return registers for function */
  49 #define REG_IRET TREG_EAX /* single word int return register */
  50 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  51 #define REG_FRET TREG_ST0 /* float return register */
  52
  53 /* defined if function parameters must be evaluated in reverse order */
  54 #define INVERT_FUNC_PARAMS
  55
  56 /* defined if structures are passed as pointers. Otherwise structures
  57    are directly pushed on stack. */
  58 //#define FUNC_STRUCT_PARAM_AS_PTR
  59
  60 /* pointer size, in bytes */
  61 #define PTR_SIZE 4
  62
  63 /* long double size and alignment, in bytes */
  64 #define LDOUBLE_SIZE  12
  65 #define LDOUBLE_ALIGN 4
  66 /* maximum alignment (for aligned attribute support) */
  67 #define MAX_ALIGN     8
  68
  69
  70 #define psym oad
  71
  72 /******************************************************/
  73 /* ELF defines */
  74
  75 #define EM_TCC_TARGET EM_386
  76
  77 /* relocation type for 32 bit data relocation */
  78 #define R_DATA_32   R_386_32
  79 #define R_DATA_PTR  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 #else /* ! TARGET_DEFS_ONLY */
  88 /******************************************************/
  89 #include "tcc.h"
  90
  91 ST_DATA const int reg_classes[NB_REGS] = {
  92     /* eax */ RC_INT | RC_EAX,
  93     /* ecx */ RC_INT | RC_ECX,
  94     /* edx */ RC_INT | RC_EDX,
  95     /* st0 */ RC_FLOAT | RC_ST0,
  96 };
  97
  98 static unsigned long func_sub_sp_offset;
  99 static int func_ret_sub;
 100 #ifdef CONFIG_TCC_BCHECK
 101 static unsigned long func_bound_offset;
 102 #endif
 103
 104 /* XXX: make it faster ? */
 105 ST_FUNC void g(int c)
 106 {
 107     int ind1;
 108     ind1 = ind + 1;
 109     if (ind1 > cur_text_section->data_allocated)
 110         section_realloc(cur_text_section, ind1);
 111     cur_text_section->data[ind] = c;
 112     ind = ind1;
 113 }
 114
 115 ST_FUNC void o(unsigned int c)
 116 {
 117     while (c) {
 118         g(c);
 119         c = c >> 8;
 120     }
 121 }
 122
 123 ST_FUNC void gen_le16(int v)
 124 {
 125     g(v);
 126     g(v >> 8);
 127 }
 128
 129 ST_FUNC void gen_le32(int c)
 130 {
 131     g(c);
 132     g(c >> 8);
 133     g(c >> 16);
 134     g(c >> 24);
 135 }
 136
 137 /* output a symbol and patch all calls to it */
 138 ST_FUNC void gsym_addr(int t, int a)
 139 {
 140     int n, *ptr;
 141     while (t) {
 142         ptr = (int *)(cur_text_section->data + t);
 143         n = *ptr; /* next value */
 144         *ptr = a - t - 4;
 145         t = n;
 146     }
 147 }
 148
 149 ST_FUNC void gsym(int t)
 150 {
 151     gsym_addr(t, ind);
 152 }
 153
 154 /* psym is used to put an instruction with a data field which is a
 155    reference to a symbol. It is in fact the same as oad ! */
 156 #define psym oad
 157
 158 /* instruction + 4 bytes data. Return the address of the data */
 159 ST_FUNC int oad(int c, int s)
 160 {
 161     int ind1;
 162
 163     o(c);
 164     ind1 = ind + 4;
 165     if (ind1 > cur_text_section->data_allocated)
 166         section_realloc(cur_text_section, ind1);
 167     *(int *)(cur_text_section->data + ind) = s;
 168     s = ind;
 169     ind = ind1;
 170     return s;
 171 }
 172
 173 /* output constant with relocation if 'r & VT_SYM' is true */
 174 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 175 {
 176     if (r & VT_SYM)
 177         greloc(cur_text_section, sym, ind, R_386_32);
 178     gen_le32(c);
 179 }
 180
 181 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 182 {
 183     if (r & VT_SYM)
 184         greloc(cur_text_section, sym, ind, R_386_PC32);
 185     gen_le32(c - 4);
 186 }
 187
 188 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 189    opcode bits */
 190 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 191 {
 192     op_reg = op_reg << 3;
 193     if ((r & VT_VALMASK) == VT_CONST) {
 194         /* constant memory reference */
 195         o(0x05 | op_reg);
 196         gen_addr32(r, sym, c);
 197     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 198         /* currently, we use only ebp as base */
 199         if (c == (char)c) {
 200             /* short reference */
 201             o(0x45 | op_reg);
 202             g(c);
 203         } else {
 204             oad(0x85 | op_reg, c);
 205         }
 206     } else {
 207         g(0x00 | op_reg | (r & VT_VALMASK));
 208     }
 209 }
 210
 211 /* load 'r' from value 'sv' */
 212 ST_FUNC void load(int r, SValue *sv)
 213 {
 214     int v, t, ft, fc, fr;
 215     SValue v1;
 216
 217 #ifdef TCC_TARGET_PE
 218     if (pe_dllimport(r, sv, load))
 219         return;
 220 #endif
 221     fr = sv->r;
 222     ft = sv->type.t;
 223     fc = sv->c.ul;
 224
 225     v = fr & VT_VALMASK;
 226     if (fr & VT_LVAL) {
 227         if (v == VT_LLOCAL) {
 228             v1.type.t = VT_INT;
 229             v1.r = VT_LOCAL | VT_LVAL;
 230             v1.c.ul = fc;
 231             load(r, &v1);
 232             fr = r;
 233         }
 234         if ((ft & VT_BTYPE) == VT_FLOAT) {
 235             o(0xd9); /* flds */
 236             r = 0;
 237         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 238             o(0xdd); /* fldl */
 239             r = 0;
 240         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 241             o(0xdb); /* fldt */
 242             r = 5;
 243         } else if ((ft & VT_TYPE) == VT_BYTE) {
 244             o(0xbe0f);   /* movsbl */
 245         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 246             o(0xb60f);   /* movzbl */
 247         } else if ((ft & VT_TYPE) == VT_SHORT) {
 248             o(0xbf0f);   /* movswl */
 249         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 250             o(0xb70f);   /* movzwl */
 251         } else {
 252             o(0x8b);     /* movl */
 253         }
 254         gen_modrm(r, fr, sv->sym, fc);
 255     } else {
 256         if (v == VT_CONST) {
 257             o(0xb8 + r); /* mov $xx, r */
 258             gen_addr32(fr, sv->sym, fc);
 259         } else if (v == VT_LOCAL) {
 260             o(0x8d); /* lea xxx(%ebp), r */
 261             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 262         } else if (v == VT_CMP) {
 263             oad(0xb8 + r, 0); /* mov $0, r */
 264             o(0x0f); /* setxx %br */
 265             o(fc);
 266             o(0xc0 + r);
 267         } else if (v == VT_JMP || v == VT_JMPI) {
 268             t = v & 1;
 269             oad(0xb8 + r, t); /* mov $1, r */
 270             o(0x05eb); /* jmp after */
 271             gsym(fc);
 272             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 273         } else if (v != r) {
 274             o(0x89);
 275             o(0xc0 + r + v * 8); /* mov v, r */
 276         }
 277     }
 278 }
 279
 280 /* store register 'r' in lvalue 'v' */
 281 ST_FUNC void store(int r, SValue *v)
 282 {
 283     int fr, bt, ft, fc;
 284
 285 #ifdef TCC_TARGET_PE
 286     if (pe_dllimport(r, v, store))
 287         return;
 288 #endif
 289     ft = v->type.t;
 290     fc = v->c.ul;
 291     fr = v->r & VT_VALMASK;
 292     bt = ft & VT_BTYPE;
 293     /* XXX: incorrect if float reg to reg */
 294     if (bt == VT_FLOAT) {
 295         o(0xd9); /* fsts */
 296         r = 2;
 297     } else if (bt == VT_DOUBLE) {
 298         o(0xdd); /* fstpl */
 299         r = 2;
 300     } else if (bt == VT_LDOUBLE) {
 301         o(0xc0d9); /* fld %st(0) */
 302         o(0xdb); /* fstpt */
 303         r = 7;
 304     } else {
 305         if (bt == VT_SHORT)
 306             o(0x66);
 307         if (bt == VT_BYTE || bt == VT_BOOL)
 308             o(0x88);
 309         else
 310             o(0x89);
 311     }
 312     if (fr == VT_CONST ||
 313         fr == VT_LOCAL ||
 314         (v->r & VT_LVAL)) {
 315         gen_modrm(r, v->r, v->sym, fc);
 316     } else if (fr != r) {
 317         o(0xc0 + fr + r * 8); /* mov r, fr */
 318     }
 319 }
 320
 321 static void gadd_sp(int val)
 322 {
 323     if (val == (char)val) {
 324         o(0xc483);
 325         g(val);
 326     } else {
 327         oad(0xc481, val); /* add $xxx, %esp */
 328     }
 329 }
 330
 331 /* 'is_jmp' is '1' if it is a jump */
 332 static void gcall_or_jmp(int is_jmp)
 333 {
 334     int r;
 335     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 336         /* constant case */
 337         if (vtop->r & VT_SYM) {
 338             /* relocation case */
 339             greloc(cur_text_section, vtop->sym,
 340                    ind + 1, R_386_PC32);
 341         } else {
 342             /* put an empty PC32 relocation */
 343             put_elf_reloc(symtab_section, cur_text_section,
 344                           ind + 1, R_386_PC32, 0);
 345         }
 346         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 347     } else {
 348         /* otherwise, indirect call */
 349         r = gv(RC_INT);
 350         o(0xff); /* call/jmp *r */
 351         o(0xd0 + r + (is_jmp << 4));
 352     }
 353 }
 354
 355 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 356 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 357
 358 /* Generate function call. The function address is pushed first, then
 359    all the parameters in call order. This functions pops all the
 360    parameters and the function address. */
 361 ST_FUNC void gfunc_call(int nb_args)
 362 {
 363     int size, align, r, args_size, i, func_call;
 364     Sym *func_sym;
 365
 366     args_size = 0;
 367     for(i = 0;i < nb_args; i++) {
 368         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 369             size = type_size(&vtop->type, &align);
 370             /* align to stack align size */
 371             size = (size + 3) & ~3;
 372             /* allocate the necessary size on stack */
 373             oad(0xec81, size); /* sub $xxx, %esp */
 374             /* generate structure store */
 375             r = get_reg(RC_INT);
 376             o(0x89); /* mov %esp, r */
 377             o(0xe0 + r);
 378             vset(&vtop->type, r | VT_LVAL, 0);
 379             vswap();
 380             vstore();
 381             args_size += size;
 382         } else if (is_float(vtop->type.t)) {
 383             gv(RC_FLOAT); /* only one float register */
 384             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 385                 size = 4;
 386             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 387                 size = 8;
 388             else
 389                 size = 12;
 390             oad(0xec81, size); /* sub $xxx, %esp */
 391             if (size == 12)
 392                 o(0x7cdb);
 393             else
 394                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 395             g(0x24);
 396             g(0x00);
 397             args_size += size;
 398         } else {
 399             /* simple type (currently always same size) */
 400             /* XXX: implicit cast ? */
 401             r = gv(RC_INT);
 402             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 403                 size = 8;
 404                 o(0x50 + vtop->r2); /* push r */
 405             } else {
 406                 size = 4;
 407             }
 408             o(0x50 + r); /* push r */
 409             args_size += size;
 410         }
 411         vtop--;
 412     }
 413     save_regs(0); /* save used temporary registers */
 414     func_sym = vtop->type.ref;
 415     func_call = FUNC_CALL(func_sym->r);
 416     /* fast call case */
 417     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 418         func_call == FUNC_FASTCALLW) {
 419         int fastcall_nb_regs;
 420         uint8_t *fastcall_regs_ptr;
 421         if (func_call == FUNC_FASTCALLW) {
 422             fastcall_regs_ptr = fastcallw_regs;
 423             fastcall_nb_regs = 2;
 424         } else {
 425             fastcall_regs_ptr = fastcall_regs;
 426             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 427         }
 428         for(i = 0;i < fastcall_nb_regs; i++) {
 429             if (args_size <= 0)
 430                 break;
 431             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 432             /* XXX: incorrect for struct/floats */
 433             args_size -= 4;
 434         }
 435     }
 436     gcall_or_jmp(0);
 437
 438 #ifdef TCC_TARGET_PE
 439     if ((func_sym->type.t & VT_BTYPE) == VT_STRUCT)
 440         args_size -= 4;
 441 #endif
 442     if (args_size && func_call != FUNC_STDCALL)
 443         gadd_sp(args_size);
 444     vtop--;
 445 }
 446
 447 #ifdef TCC_TARGET_PE
 448 #define FUNC_PROLOG_SIZE 10
 449 #else
 450 #define FUNC_PROLOG_SIZE 9
 451 #endif
 452
 453 /* generate function prolog of type 't' */
 454 ST_FUNC void gfunc_prolog(CType *func_type)
 455 {
 456     int addr, align, size, func_call, fastcall_nb_regs;
 457     int param_index, param_addr;
 458     uint8_t *fastcall_regs_ptr;
 459     Sym *sym;
 460     CType *type;
 461
 462     sym = func_type->ref;
 463     func_call = FUNC_CALL(sym->r);
 464     addr = 8;
 465     loc = 0;
 466     func_vc = 0;
 467
 468     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 469         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 470         fastcall_regs_ptr = fastcall_regs;
 471     } else if (func_call == FUNC_FASTCALLW) {
 472         fastcall_nb_regs = 2;
 473         fastcall_regs_ptr = fastcallw_regs;
 474     } else {
 475         fastcall_nb_regs = 0;
 476         fastcall_regs_ptr = NULL;
 477     }
 478     param_index = 0;
 479
 480     ind += FUNC_PROLOG_SIZE;
 481     func_sub_sp_offset = ind;
 482     /* if the function returns a structure, then add an
 483        implicit pointer parameter */
 484     func_vt = sym->type;
 485     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 486         /* XXX: fastcall case ? */
 487         func_vc = addr;
 488         addr += 4;
 489         param_index++;
 490     }
 491     /* define parameters */
 492     while ((sym = sym->next) != NULL) {
 493         type = &sym->type;
 494         size = type_size(type, &align);
 495         size = (size + 3) & ~3;
 496 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 497         /* structs are passed as pointer */
 498         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 499             size = 4;
 500         }
 501 #endif
 502         if (param_index < fastcall_nb_regs) {
 503             /* save FASTCALL register */
 504             loc -= 4;
 505             o(0x89);     /* movl */
 506             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 507             param_addr = loc;
 508         } else {
 509             param_addr = addr;
 510             addr += size;
 511         }
 512         sym_push(sym->v & ~SYM_FIELD, type,
 513                  VT_LOCAL | lvalue_type(type->t), param_addr);
 514         param_index++;
 515     }
 516     func_ret_sub = 0;
 517     /* pascal type call ? */
 518     if (func_call == FUNC_STDCALL)
 519         func_ret_sub = addr - 8;
 520 #ifdef TCC_TARGET_PE
 521     else if (func_vc)
 522         func_ret_sub = 4;
 523 #endif
 524
 525 #ifdef CONFIG_TCC_BCHECK
 526     /* leave some room for bound checking code */
 527     if (tcc_state->do_bounds_check) {
 528         oad(0xb8, 0); /* lbound section pointer */
 529         oad(0xb8, 0); /* call to function */
 530         func_bound_offset = lbounds_section->data_offset;
 531     }
 532 #endif
 533 }
 534
 535 /* generate function epilog */
 536 ST_FUNC void gfunc_epilog(void)
 537 {
 538     int v, saved_ind;
 539
 540 #ifdef CONFIG_TCC_BCHECK
 541     if (tcc_state->do_bounds_check
 542      && func_bound_offset != lbounds_section->data_offset) {
 543         int saved_ind;
 544         int *bounds_ptr;
 545         Sym *sym, *sym_data;
 546         /* add end of table info */
 547         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 548         *bounds_ptr = 0;
 549         /* generate bound local allocation */
 550         saved_ind = ind;
 551         ind = func_sub_sp_offset;
 552         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 553                                func_bound_offset, lbounds_section->data_offset);
 554         greloc(cur_text_section, sym_data,
 555                ind + 1, R_386_32);
 556         oad(0xb8, 0); /* mov %eax, xxx */
 557         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 558         greloc(cur_text_section, sym,
 559                ind + 1, R_386_PC32);
 560         oad(0xe8, -4);
 561         ind = saved_ind;
 562         /* generate bound check local freeing */
 563         o(0x5250); /* save returned value, if any */
 564         greloc(cur_text_section, sym_data,
 565                ind + 1, R_386_32);
 566         oad(0xb8, 0); /* mov %eax, xxx */
 567         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 568         greloc(cur_text_section, sym,
 569                ind + 1, R_386_PC32);
 570         oad(0xe8, -4);
 571         o(0x585a); /* restore returned value, if any */
 572     }
 573 #endif
 574     o(0xc9); /* leave */
 575     if (func_ret_sub == 0) {
 576         o(0xc3); /* ret */
 577     } else {
 578         o(0xc2); /* ret n */
 579         g(func_ret_sub);
 580         g(func_ret_sub >> 8);
 581     }
 582     /* align local size to word & save local variables */
 583
 584     v = (-loc + 3) & -4;
 585     saved_ind = ind;
 586     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 587 #ifdef TCC_TARGET_PE
 588     if (v >= 4096) {
 589         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 590         oad(0xb8, v); /* mov stacksize, %eax */
 591         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 592         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 593     } else
 594 #endif
 595     {
 596         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 597         o(0xec81);  /* sub esp, stacksize */
 598         gen_le32(v);
 599 #if FUNC_PROLOG_SIZE == 10
 600         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 601 #endif
 602     }
 603     ind = saved_ind;
 604 }
 605
 606 /* generate a jump to a label */
 607 ST_FUNC int gjmp(int t)
 608 {
 609     return psym(0xe9, t);
 610 }
 611
 612 /* generate a jump to a fixed address */
 613 ST_FUNC void gjmp_addr(int a)
 614 {
 615     int r;
 616     r = a - ind - 2;
 617     if (r == (char)r) {
 618         g(0xeb);
 619         g(r);
 620     } else {
 621         oad(0xe9, a - ind - 5);
 622     }
 623 }
 624
 625 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 626 ST_FUNC int gtst(int inv, int t)
 627 {
 628     int v, *p;
 629
 630     v = vtop->r & VT_VALMASK;
 631     if (v == VT_CMP) {
 632         /* fast case : can jump directly since flags are set */
 633         g(0x0f);
 634         t = psym((vtop->c.i - 16) ^ inv, t);
 635     } else if (v == VT_JMP || v == VT_JMPI) {
 636         /* && or || optimization */
 637         if ((v & 1) == inv) {
 638             /* insert vtop->c jump list in t */
 639             p = &vtop->c.i;
 640             while (*p != 0)
 641                 p = (int *)(cur_text_section->data + *p);
 642             *p = t;
 643             t = vtop->c.i;
 644         } else {
 645             t = gjmp(t);
 646             gsym(vtop->c.i);
 647         }
 648     } else {
 649         if (is_float(vtop->type.t) ||
 650             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 651             vpushi(0);
 652             gen_op(TOK_NE);
 653         }
 654         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 655             /* constant jmp optimization */
 656             if ((vtop->c.i != 0) != inv)
 657                 t = gjmp(t);
 658         } else {
 659             v = gv(RC_INT);
 660             o(0x85);
 661             o(0xc0 + v * 9);
 662             g(0x0f);
 663             t = psym(0x85 ^ inv, t);
 664         }
 665     }
 666     vtop--;
 667     return t;
 668 }
 669
 670 /* generate an integer binary operation */
 671 ST_FUNC void gen_opi(int op)
 672 {
 673     int r, fr, opc, c;
 674
 675     switch(op) {
 676     case '+':
 677     case TOK_ADDC1: /* add with carry generation */
 678         opc = 0;
 679     gen_op8:
 680         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 681             /* constant case */
 682             vswap();
 683             r = gv(RC_INT);
 684             vswap();
 685             c = vtop->c.i;
 686             if (c == (char)c) {
 687                 /* XXX: generate inc and dec for smaller code ? */
 688                 o(0x83);
 689                 o(0xc0 | (opc << 3) | r);
 690                 g(c);
 691             } else {
 692                 o(0x81);
 693                 oad(0xc0 | (opc << 3) | r, c);
 694             }
 695         } else {
 696             gv2(RC_INT, RC_INT);
 697             r = vtop[-1].r;
 698             fr = vtop[0].r;
 699             o((opc << 3) | 0x01);
 700             o(0xc0 + r + fr * 8);
 701         }
 702         vtop--;
 703         if (op >= TOK_ULT && op <= TOK_GT) {
 704             vtop->r = VT_CMP;
 705             vtop->c.i = op;
 706         }
 707         break;
 708     case '-':
 709     case TOK_SUBC1: /* sub with carry generation */
 710         opc = 5;
 711         goto gen_op8;
 712     case TOK_ADDC2: /* add with carry use */
 713         opc = 2;
 714         goto gen_op8;
 715     case TOK_SUBC2: /* sub with carry use */
 716         opc = 3;
 717         goto gen_op8;
 718     case '&':
 719         opc = 4;
 720         goto gen_op8;
 721     case '^':
 722         opc = 6;
 723         goto gen_op8;
 724     case '|':
 725         opc = 1;
 726         goto gen_op8;
 727     case '*':
 728         gv2(RC_INT, RC_INT);
 729         r = vtop[-1].r;
 730         fr = vtop[0].r;
 731         vtop--;
 732         o(0xaf0f); /* imul fr, r */
 733         o(0xc0 + fr + r * 8);
 734         break;
 735     case TOK_SHL:
 736         opc = 4;
 737         goto gen_shift;
 738     case TOK_SHR:
 739         opc = 5;
 740         goto gen_shift;
 741     case TOK_SAR:
 742         opc = 7;
 743     gen_shift:
 744         opc = 0xc0 | (opc << 3);
 745         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 746             /* constant case */
 747             vswap();
 748             r = gv(RC_INT);
 749             vswap();
 750             c = vtop->c.i & 0x1f;
 751             o(0xc1); /* shl/shr/sar $xxx, r */
 752             o(opc | r);
 753             g(c);
 754         } else {
 755             /* we generate the shift in ecx */
 756             gv2(RC_INT, RC_ECX);
 757             r = vtop[-1].r;
 758             o(0xd3); /* shl/shr/sar %cl, r */
 759             o(opc | r);
 760         }
 761         vtop--;
 762         break;
 763     case '/':
 764     case TOK_UDIV:
 765     case TOK_PDIV:
 766     case '%':
 767     case TOK_UMOD:
 768     case TOK_UMULL:
 769         /* first operand must be in eax */
 770         /* XXX: need better constraint for second operand */
 771         gv2(RC_EAX, RC_ECX);
 772         r = vtop[-1].r;
 773         fr = vtop[0].r;
 774         vtop--;
 775         save_reg(TREG_EDX);
 776         if (op == TOK_UMULL) {
 777             o(0xf7); /* mul fr */
 778             o(0xe0 + fr);
 779             vtop->r2 = TREG_EDX;
 780             r = TREG_EAX;
 781         } else {
 782             if (op == TOK_UDIV || op == TOK_UMOD) {
 783                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 784                 o(0xf0 + fr);
 785             } else {
 786                 o(0xf799); /* cltd, idiv fr, %eax */
 787                 o(0xf8 + fr);
 788             }
 789             if (op == '%' || op == TOK_UMOD)
 790                 r = TREG_EDX;
 791             else
 792                 r = TREG_EAX;
 793         }
 794         vtop->r = r;
 795         break;
 796     default:
 797         opc = 7;
 798         goto gen_op8;
 799     }
 800 }
 801
 802 /* generate a floating point operation 'v = t1 op t2' instruction. The
 803    two operands are guaranted to have the same floating point type */
 804 /* XXX: need to use ST1 too */
 805 ST_FUNC void gen_opf(int op)
 806 {
 807     int a, ft, fc, swapped, r;
 808
 809     /* convert constants to memory references */
 810     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 811         vswap();
 812         gv(RC_FLOAT);
 813         vswap();
 814     }
 815     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 816         gv(RC_FLOAT);
 817
 818     /* must put at least one value in the floating point register */
 819     if ((vtop[-1].r & VT_LVAL) &&
 820         (vtop[0].r & VT_LVAL)) {
 821         vswap();
 822         gv(RC_FLOAT);
 823         vswap();
 824     }
 825     swapped = 0;
 826     /* swap the stack if needed so that t1 is the register and t2 is
 827        the memory reference */
 828     if (vtop[-1].r & VT_LVAL) {
 829         vswap();
 830         swapped = 1;
 831     }
 832     if (op >= TOK_ULT && op <= TOK_GT) {
 833         /* load on stack second operand */
 834         load(TREG_ST0, vtop);
 835         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 836         if (op == TOK_GE || op == TOK_GT)
 837             swapped = !swapped;
 838         else if (op == TOK_EQ || op == TOK_NE)
 839             swapped = 0;
 840         if (swapped)
 841             o(0xc9d9); /* fxch %st(1) */
 842         o(0xe9da); /* fucompp */
 843         o(0xe0df); /* fnstsw %ax */
 844         if (op == TOK_EQ) {
 845             o(0x45e480); /* and $0x45, %ah */
 846             o(0x40fC80); /* cmp $0x40, %ah */
 847         } else if (op == TOK_NE) {
 848             o(0x45e480); /* and $0x45, %ah */
 849             o(0x40f480); /* xor $0x40, %ah */
 850             op = TOK_NE;
 851         } else if (op == TOK_GE || op == TOK_LE) {
 852             o(0x05c4f6); /* test $0x05, %ah */
 853             op = TOK_EQ;
 854         } else {
 855             o(0x45c4f6); /* test $0x45, %ah */
 856             op = TOK_EQ;
 857         }
 858         vtop--;
 859         vtop->r = VT_CMP;
 860         vtop->c.i = op;
 861     } else {
 862         /* no memory reference possible for long double operations */
 863         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 864             load(TREG_ST0, vtop);
 865             swapped = !swapped;
 866         }
 867
 868         switch(op) {
 869         default:
 870         case '+':
 871             a = 0;
 872             break;
 873         case '-':
 874             a = 4;
 875             if (swapped)
 876                 a++;
 877             break;
 878         case '*':
 879             a = 1;
 880             break;
 881         case '/':
 882             a = 6;
 883             if (swapped)
 884                 a++;
 885             break;
 886         }
 887         ft = vtop->type.t;
 888         fc = vtop->c.ul;
 889         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 890             o(0xde); /* fxxxp %st, %st(1) */
 891             o(0xc1 + (a << 3));
 892         } else {
 893             /* if saved lvalue, then we must reload it */
 894             r = vtop->r;
 895             if ((r & VT_VALMASK) == VT_LLOCAL) {
 896                 SValue v1;
 897                 r = get_reg(RC_INT);
 898                 v1.type.t = VT_INT;
 899                 v1.r = VT_LOCAL | VT_LVAL;
 900                 v1.c.ul = fc;
 901                 load(r, &v1);
 902                 fc = 0;
 903             }
 904
 905             if ((ft & VT_BTYPE) == VT_DOUBLE)
 906                 o(0xdc);
 907             else
 908                 o(0xd8);
 909             gen_modrm(a, r, vtop->sym, fc);
 910         }
 911         vtop--;
 912     }
 913 }
 914
 915 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 916    and 'long long' cases. */
 917 ST_FUNC void gen_cvt_itof(int t)
 918 {
 919     save_reg(TREG_ST0);
 920     gv(RC_INT);
 921     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 922         /* signed long long to float/double/long double (unsigned case
 923            is handled generically) */
 924         o(0x50 + vtop->r2); /* push r2 */
 925         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 926         o(0x242cdf); /* fildll (%esp) */
 927         o(0x08c483); /* add $8, %esp */
 928     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 929                (VT_INT | VT_UNSIGNED)) {
 930         /* unsigned int to float/double/long double */
 931         o(0x6a); /* push $0 */
 932         g(0x00);
 933         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 934         o(0x242cdf); /* fildll (%esp) */
 935         o(0x08c483); /* add $8, %esp */
 936     } else {
 937         /* int to float/double/long double */
 938         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 939         o(0x2404db); /* fildl (%esp) */
 940         o(0x04c483); /* add $4, %esp */
 941     }
 942     vtop->r = TREG_ST0;
 943 }
 944
 945 /* convert fp to int 't' type */
 946 /* XXX: handle long long case */
 947 ST_FUNC void gen_cvt_ftoi(int t)
 948 {
 949     int r, r2, size;
 950     Sym *sym;
 951     CType ushort_type;
 952
 953     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 954     ushort_type.ref = 0;
 955
 956     gv(RC_FLOAT);
 957     if (t != VT_INT)
 958         size = 8;
 959     else
 960         size = 4;
 961
 962     o(0x2dd9); /* ldcw xxx */
 963     sym = external_global_sym(TOK___tcc_int_fpu_control,
 964                               &ushort_type, VT_LVAL);
 965     greloc(cur_text_section, sym,
 966            ind, R_386_32);
 967     gen_le32(0);
 968
 969     oad(0xec81, size); /* sub $xxx, %esp */
 970     if (size == 4)
 971         o(0x1cdb); /* fistpl */
 972     else
 973         o(0x3cdf); /* fistpll */
 974     o(0x24);
 975     o(0x2dd9); /* ldcw xxx */
 976     sym = external_global_sym(TOK___tcc_fpu_control,
 977                               &ushort_type, VT_LVAL);
 978     greloc(cur_text_section, sym,
 979            ind, R_386_32);
 980     gen_le32(0);
 981
 982     r = get_reg(RC_INT);
 983     o(0x58 + r); /* pop r */
 984     if (size == 8) {
 985         if (t == VT_LLONG) {
 986             vtop->r = r; /* mark reg as used */
 987             r2 = get_reg(RC_INT);
 988             o(0x58 + r2); /* pop r2 */
 989             vtop->r2 = r2;
 990         } else {
 991             o(0x04c483); /* add $4, %esp */
 992         }
 993     }
 994     vtop->r = r;
 995 }
 996
 997 /* convert from one floating point type to another */
 998 ST_FUNC void gen_cvt_ftof(int t)
 999 {
1000     /* all we have to do on i386 is to put the float in a register */
1001     gv(RC_FLOAT);
1002 }
1003
1004 /* computed goto support */
1005 ST_FUNC void ggoto(void)
1006 {
1007     gcall_or_jmp(1);
1008     vtop--;
1009 }
1010
1011 /* bound check support functions */
1012 #ifdef CONFIG_TCC_BCHECK
1013
1014 /* generate a bounded pointer addition */
1015 ST_FUNC void gen_bounded_ptr_add(void)
1016 {
1017     Sym *sym;
1018
1019     /* prepare fast i386 function call (args in eax and edx) */
1020     gv2(RC_EAX, RC_EDX);
1021     /* save all temporary registers */
1022     vtop -= 2;
1023     save_regs(0);
1024     /* do a fast function call */
1025     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
1026     greloc(cur_text_section, sym,
1027            ind + 1, R_386_PC32);
1028     oad(0xe8, -4);
1029     /* returned pointer is in eax */
1030     vtop++;
1031     vtop->r = TREG_EAX | VT_BOUNDED;
1032     /* address of bounding function call point */
1033     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1034 }
1035
1036 /* patch pointer addition in vtop so that pointer dereferencing is
1037    also tested */
1038 ST_FUNC void gen_bounded_ptr_deref(void)
1039 {
1040     int func;
1041     int size, align;
1042     Elf32_Rel *rel;
1043     Sym *sym;
1044
1045     size = 0;
1046     /* XXX: put that code in generic part of tcc */
1047     if (!is_float(vtop->type.t)) {
1048         if (vtop->r & VT_LVAL_BYTE)
1049             size = 1;
1050         else if (vtop->r & VT_LVAL_SHORT)
1051             size = 2;
1052     }
1053     if (!size)
1054         size = type_size(&vtop->type, &align);
1055     switch(size) {
1056     case  1: func = TOK___bound_ptr_indir1; break;
1057     case  2: func = TOK___bound_ptr_indir2; break;
1058     case  4: func = TOK___bound_ptr_indir4; break;
1059     case  8: func = TOK___bound_ptr_indir8; break;
1060     case 12: func = TOK___bound_ptr_indir12; break;
1061     case 16: func = TOK___bound_ptr_indir16; break;
1062     default:
1063         error("unhandled size when derefencing bounded pointer");
1064         func = 0;
1065         break;
1066     }
1067
1068     /* patch relocation */
1069     /* XXX: find a better solution ? */
1070     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1071     sym = external_global_sym(func, &func_old_type, 0);
1072     if (!sym->c)
1073         put_extern_sym(sym, NULL, 0, 0);
1074     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1075 }
1076 #endif
1077
1078 /* end of X86 code generator */
1079 /*************************************************************/
1080 #endif
1081 /*************************************************************/