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 #ifdef TCC_TARGET_PE
 188 static void mk_pointer(CType *type);
 189 static void indir(void);
 190
 191 int handle_dllimport(int r, SValue *sv, void (*fn)(int r, SValue *sv))
 192 {
 193     if ((sv->r & (VT_VALMASK|VT_SYM|VT_CONST)) != (VT_SYM|VT_CONST))
 194         return 0;
 195     if (0 == (sv->sym->type.t & VT_IMPORT))
 196         return 0;
 197
 198     printf("import %d %04x %s\n", r, ind, get_tok_str(sv->sym->v, NULL));
 199
 200     sv->sym->type.t &= ~VT_IMPORT;
 201     ++vtop;
 202
 203     *vtop = *sv;
 204     mk_pointer(&vtop->type);
 205     indir();
 206     fn(r, vtop);
 207
 208     --vtop;
 209     sv->sym->type.t |= VT_IMPORT;
 210     return 1;
 211 }
 212 #endif
 213
 214 /* load 'r' from value 'sv' */
 215 void load(int r, SValue *sv)
 216 {
 217     int v, t, ft, fc, fr;
 218     SValue v1;
 219
 220 #ifdef TCC_TARGET_PE
 221     if (handle_dllimport(r, sv, load))
 222         return;
 223 #endif
 224     fr = sv->r;
 225     ft = sv->type.t;
 226     fc = sv->c.ul;
 227
 228     v = fr & VT_VALMASK;
 229     if (fr & VT_LVAL) {
 230         if (v == VT_LLOCAL) {
 231             v1.type.t = VT_INT;
 232             v1.r = VT_LOCAL | VT_LVAL;
 233             v1.c.ul = fc;
 234             load(r, &v1);
 235             fr = r;
 236         }
 237         if ((ft & VT_BTYPE) == VT_FLOAT) {
 238             o(0xd9); /* flds */
 239             r = 0;
 240         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 241             o(0xdd); /* fldl */
 242             r = 0;
 243         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 244             o(0xdb); /* fldt */
 245             r = 5;
 246         } else if ((ft & VT_TYPE) == VT_BYTE) {
 247             o(0xbe0f);   /* movsbl */
 248         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 249             o(0xb60f);   /* movzbl */
 250         } else if ((ft & VT_TYPE) == VT_SHORT) {
 251             o(0xbf0f);   /* movswl */
 252         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 253             o(0xb70f);   /* movzwl */
 254         } else {
 255             o(0x8b);     /* movl */
 256         }
 257         gen_modrm(r, fr, sv->sym, fc);
 258     } else {
 259         if (v == VT_CONST) {
 260             o(0xb8 + r); /* mov $xx, r */
 261             gen_addr32(fr, sv->sym, fc);
 262         } else if (v == VT_LOCAL) {
 263             o(0x8d); /* lea xxx(%ebp), r */
 264             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 265         } else if (v == VT_CMP) {
 266             oad(0xb8 + r, 0); /* mov $0, r */
 267             o(0x0f); /* setxx %br */
 268             o(fc);
 269             o(0xc0 + r);
 270         } else if (v == VT_JMP || v == VT_JMPI) {
 271             t = v & 1;
 272             oad(0xb8 + r, t); /* mov $1, r */
 273             o(0x05eb); /* jmp after */
 274             gsym(fc);
 275             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 276         } else if (v != r) {
 277             o(0x89);
 278             o(0xc0 + r + v * 8); /* mov v, r */
 279         }
 280     }
 281 }
 282
 283 /* store register 'r' in lvalue 'v' */
 284 void store(int r, SValue *v)
 285 {
 286     int fr, bt, ft, fc;
 287
 288 #ifdef TCC_TARGET_PE
 289     if (handle_dllimport(r, v, store))
 290         return;
 291 #endif
 292     ft = v->type.t;
 293     fc = v->c.ul;
 294     fr = v->r & VT_VALMASK;
 295     bt = ft & VT_BTYPE;
 296     /* XXX: incorrect if float reg to reg */
 297     if (bt == VT_FLOAT) {
 298         o(0xd9); /* fsts */
 299         r = 2;
 300     } else if (bt == VT_DOUBLE) {
 301         o(0xdd); /* fstpl */
 302         r = 2;
 303     } else if (bt == VT_LDOUBLE) {
 304         o(0xc0d9); /* fld %st(0) */
 305         o(0xdb); /* fstpt */
 306         r = 7;
 307     } else {
 308         if (bt == VT_SHORT)
 309             o(0x66);
 310         if (bt == VT_BYTE || bt == VT_BOOL)
 311             o(0x88);
 312         else
 313             o(0x89);
 314     }
 315     if (fr == VT_CONST ||
 316         fr == VT_LOCAL ||
 317         (v->r & VT_LVAL)) {
 318         gen_modrm(r, v->r, v->sym, fc);
 319     } else if (fr != r) {
 320         o(0xc0 + fr + r * 8); /* mov r, fr */
 321     }
 322 }
 323
 324 static void gadd_sp(int val)
 325 {
 326     if (val == (char)val) {
 327         o(0xc483);
 328         g(val);
 329     } else {
 330         oad(0xc481, val); /* add $xxx, %esp */
 331     }
 332 }
 333
 334 /* 'is_jmp' is '1' if it is a jump */
 335 static void gcall_or_jmp(int is_jmp)
 336 {
 337     int r;
 338     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 339         /* constant case */
 340         if (vtop->r & VT_SYM) {
 341             /* relocation case */
 342             greloc(cur_text_section, vtop->sym,
 343                    ind + 1, R_386_PC32);
 344         } else {
 345             /* put an empty PC32 relocation */
 346             put_elf_reloc(symtab_section, cur_text_section,
 347                           ind + 1, R_386_PC32, 0);
 348         }
 349         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 350     } else {
 351         /* otherwise, indirect call */
 352         r = gv(RC_INT);
 353         o(0xff); /* call/jmp *r */
 354         o(0xd0 + r + (is_jmp << 4));
 355     }
 356 }
 357
 358 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 359 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 360
 361 /* Generate function call. The function address is pushed first, then
 362    all the parameters in call order. This functions pops all the
 363    parameters and the function address. */
 364 void gfunc_call(int nb_args)
 365 {
 366     int size, align, r, args_size, i, func_call;
 367     Sym *func_sym;
 368
 369     args_size = 0;
 370     for(i = 0;i < nb_args; i++) {
 371         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 372             size = type_size(&vtop->type, &align);
 373             /* align to stack align size */
 374             size = (size + 3) & ~3;
 375             /* allocate the necessary size on stack */
 376             oad(0xec81, size); /* sub $xxx, %esp */
 377             /* generate structure store */
 378             r = get_reg(RC_INT);
 379             o(0x89); /* mov %esp, r */
 380             o(0xe0 + r);
 381             vset(&vtop->type, r | VT_LVAL, 0);
 382             vswap();
 383             vstore();
 384             args_size += size;
 385         } else if (is_float(vtop->type.t)) {
 386             gv(RC_FLOAT); /* only one float register */
 387             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 388                 size = 4;
 389             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 390                 size = 8;
 391             else
 392                 size = 12;
 393             oad(0xec81, size); /* sub $xxx, %esp */
 394             if (size == 12)
 395                 o(0x7cdb);
 396             else
 397                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 398             g(0x24);
 399             g(0x00);
 400             args_size += size;
 401         } else {
 402             /* simple type (currently always same size) */
 403             /* XXX: implicit cast ? */
 404             r = gv(RC_INT);
 405             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 406                 size = 8;
 407                 o(0x50 + vtop->r2); /* push r */
 408             } else {
 409                 size = 4;
 410             }
 411             o(0x50 + r); /* push r */
 412             args_size += size;
 413         }
 414         vtop--;
 415     }
 416     save_regs(0); /* save used temporary registers */
 417     func_sym = vtop->type.ref;
 418     func_call = FUNC_CALL(func_sym->r);
 419     /* fast call case */
 420     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 421         func_call == FUNC_FASTCALLW) {
 422         int fastcall_nb_regs;
 423         uint8_t *fastcall_regs_ptr;
 424         if (func_call == FUNC_FASTCALLW) {
 425             fastcall_regs_ptr = fastcallw_regs;
 426             fastcall_nb_regs = 2;
 427         } else {
 428             fastcall_regs_ptr = fastcall_regs;
 429             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 430         }
 431         for(i = 0;i < fastcall_nb_regs; i++) {
 432             if (args_size <= 0)
 433                 break;
 434             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 435             /* XXX: incorrect for struct/floats */
 436             args_size -= 4;
 437         }
 438     }
 439     gcall_or_jmp(0);
 440
 441 #ifdef TCC_TARGET_PE
 442     if ((func_sym->type.t & VT_BTYPE) == VT_STRUCT)
 443         args_size -= 4;
 444 #endif
 445     if (args_size && func_call != FUNC_STDCALL)
 446         gadd_sp(args_size);
 447     vtop--;
 448 }
 449
 450 #ifdef TCC_TARGET_PE
 451 #define FUNC_PROLOG_SIZE 10
 452 #else
 453 #define FUNC_PROLOG_SIZE 9
 454 #endif
 455
 456 /* generate function prolog of type 't' */
 457 void gfunc_prolog(CType *func_type)
 458 {
 459     int addr, align, size, func_call, fastcall_nb_regs;
 460     int param_index, param_addr;
 461     uint8_t *fastcall_regs_ptr;
 462     Sym *sym;
 463     CType *type;
 464
 465     sym = func_type->ref;
 466     func_call = FUNC_CALL(sym->r);
 467     addr = 8;
 468     loc = 0;
 469     func_vc = 0;
 470
 471     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 472         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 473         fastcall_regs_ptr = fastcall_regs;
 474     } else if (func_call == FUNC_FASTCALLW) {
 475         fastcall_nb_regs = 2;
 476         fastcall_regs_ptr = fastcallw_regs;
 477     } else {
 478         fastcall_nb_regs = 0;
 479         fastcall_regs_ptr = NULL;
 480     }
 481     param_index = 0;
 482
 483     ind += FUNC_PROLOG_SIZE;
 484     func_sub_sp_offset = ind;
 485     /* if the function returns a structure, then add an
 486        implicit pointer parameter */
 487     func_vt = sym->type;
 488     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 489         /* XXX: fastcall case ? */
 490         func_vc = addr;
 491         addr += 4;
 492         param_index++;
 493     }
 494     /* define parameters */
 495     while ((sym = sym->next) != NULL) {
 496         type = &sym->type;
 497         size = type_size(type, &align);
 498         size = (size + 3) & ~3;
 499 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 500         /* structs are passed as pointer */
 501         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 502             size = 4;
 503         }
 504 #endif
 505         if (param_index < fastcall_nb_regs) {
 506             /* save FASTCALL register */
 507             loc -= 4;
 508             o(0x89);     /* movl */
 509             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 510             param_addr = loc;
 511         } else {
 512             param_addr = addr;
 513             addr += size;
 514         }
 515         sym_push(sym->v & ~SYM_FIELD, type,
 516                  VT_LOCAL | lvalue_type(type->t), param_addr);
 517         param_index++;
 518     }
 519     func_ret_sub = 0;
 520     /* pascal type call ? */
 521     if (func_call == FUNC_STDCALL)
 522         func_ret_sub = addr - 8;
 523 #ifdef TCC_TARGET_PE
 524     else if (func_vc)
 525         func_ret_sub = 4;
 526 #endif
 527
 528     /* leave some room for bound checking code */
 529     if (tcc_state->do_bounds_check) {
 530         oad(0xb8, 0); /* lbound section pointer */
 531         oad(0xb8, 0); /* call to function */
 532         func_bound_offset = lbounds_section->data_offset;
 533     }
 534 }
 535
 536 /* generate function epilog */
 537 void gfunc_epilog(void)
 538 {
 539     int v, saved_ind;
 540
 541 #ifdef CONFIG_TCC_BCHECK
 542     if (tcc_state->do_bounds_check
 543      && func_bound_offset != lbounds_section->data_offset) {
 544         int saved_ind;
 545         int *bounds_ptr;
 546         Sym *sym, *sym_data;
 547         /* add end of table info */
 548         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 549         *bounds_ptr = 0;
 550         /* generate bound local allocation */
 551         saved_ind = ind;
 552         ind = func_sub_sp_offset;
 553         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 554                                func_bound_offset, lbounds_section->data_offset);
 555         greloc(cur_text_section, sym_data,
 556                ind + 1, R_386_32);
 557         oad(0xb8, 0); /* mov %eax, xxx */
 558         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 559         greloc(cur_text_section, sym,
 560                ind + 1, R_386_PC32);
 561         oad(0xe8, -4);
 562         ind = saved_ind;
 563         /* generate bound check local freeing */
 564         o(0x5250); /* save returned value, if any */
 565         greloc(cur_text_section, sym_data,
 566                ind + 1, R_386_32);
 567         oad(0xb8, 0); /* mov %eax, xxx */
 568         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 569         greloc(cur_text_section, sym,
 570                ind + 1, R_386_PC32);
 571         oad(0xe8, -4);
 572         o(0x585a); /* restore returned value, if any */
 573     }
 574 #endif
 575     o(0xc9); /* leave */
 576     if (func_ret_sub == 0) {
 577         o(0xc3); /* ret */
 578     } else {
 579         o(0xc2); /* ret n */
 580         g(func_ret_sub);
 581         g(func_ret_sub >> 8);
 582     }
 583     /* align local size to word & save local variables */
 584
 585     v = (-loc + 3) & -4;
 586     saved_ind = ind;
 587     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 588 #ifdef TCC_TARGET_PE
 589     if (v >= 4096) {
 590         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 591         oad(0xb8, v); /* mov stacksize, %eax */
 592         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 593         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 594     } else
 595 #endif
 596     {
 597         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 598         o(0xec81);  /* sub esp, stacksize */
 599         gen_le32(v);
 600 #if FUNC_PROLOG_SIZE == 10
 601         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 602 #endif
 603     }
 604     ind = saved_ind;
 605 }
 606
 607 /* generate a jump to a label */
 608 int gjmp(int t)
 609 {
 610     return psym(0xe9, t);
 611 }
 612
 613 /* generate a jump to a fixed address */
 614 void gjmp_addr(int a)
 615 {
 616     int r;
 617     r = a - ind - 2;
 618     if (r == (char)r) {
 619         g(0xeb);
 620         g(r);
 621     } else {
 622         oad(0xe9, a - ind - 5);
 623     }
 624 }
 625
 626 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 627 int gtst(int inv, int t)
 628 {
 629     int v, *p;
 630
 631     v = vtop->r & VT_VALMASK;
 632     if (v == VT_CMP) {
 633         /* fast case : can jump directly since flags are set */
 634         g(0x0f);
 635         t = psym((vtop->c.i - 16) ^ inv, t);
 636     } else if (v == VT_JMP || v == VT_JMPI) {
 637         /* && or || optimization */
 638         if ((v & 1) == inv) {
 639             /* insert vtop->c jump list in t */
 640             p = &vtop->c.i;
 641             while (*p != 0)
 642                 p = (int *)(cur_text_section->data + *p);
 643             *p = t;
 644             t = vtop->c.i;
 645         } else {
 646             t = gjmp(t);
 647             gsym(vtop->c.i);
 648         }
 649     } else {
 650         if (is_float(vtop->type.t) ||
 651             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 652             vpushi(0);
 653             gen_op(TOK_NE);
 654         }
 655         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 656             /* constant jmp optimization */
 657             if ((vtop->c.i != 0) != inv)
 658                 t = gjmp(t);
 659         } else {
 660             v = gv(RC_INT);
 661             o(0x85);
 662             o(0xc0 + v * 9);
 663             g(0x0f);
 664             t = psym(0x85 ^ inv, t);
 665         }
 666     }
 667     vtop--;
 668     return t;
 669 }
 670
 671 /* generate an integer binary operation */
 672 void gen_opi(int op)
 673 {
 674     int r, fr, opc, c;
 675
 676     switch(op) {
 677     case '+':
 678     case TOK_ADDC1: /* add with carry generation */
 679         opc = 0;
 680     gen_op8:
 681         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 682             /* constant case */
 683             vswap();
 684             r = gv(RC_INT);
 685             vswap();
 686             c = vtop->c.i;
 687             if (c == (char)c) {
 688                 /* XXX: generate inc and dec for smaller code ? */
 689                 o(0x83);
 690                 o(0xc0 | (opc << 3) | r);
 691                 g(c);
 692             } else {
 693                 o(0x81);
 694                 oad(0xc0 | (opc << 3) | r, c);
 695             }
 696         } else {
 697             gv2(RC_INT, RC_INT);
 698             r = vtop[-1].r;
 699             fr = vtop[0].r;
 700             o((opc << 3) | 0x01);
 701             o(0xc0 + r + fr * 8);
 702         }
 703         vtop--;
 704         if (op >= TOK_ULT && op <= TOK_GT) {
 705             vtop->r = VT_CMP;
 706             vtop->c.i = op;
 707         }
 708         break;
 709     case '-':
 710     case TOK_SUBC1: /* sub with carry generation */
 711         opc = 5;
 712         goto gen_op8;
 713     case TOK_ADDC2: /* add with carry use */
 714         opc = 2;
 715         goto gen_op8;
 716     case TOK_SUBC2: /* sub with carry use */
 717         opc = 3;
 718         goto gen_op8;
 719     case '&':
 720         opc = 4;
 721         goto gen_op8;
 722     case '^':
 723         opc = 6;
 724         goto gen_op8;
 725     case '|':
 726         opc = 1;
 727         goto gen_op8;
 728     case '*':
 729         gv2(RC_INT, RC_INT);
 730         r = vtop[-1].r;
 731         fr = vtop[0].r;
 732         vtop--;
 733         o(0xaf0f); /* imul fr, r */
 734         o(0xc0 + fr + r * 8);
 735         break;
 736     case TOK_SHL:
 737         opc = 4;
 738         goto gen_shift;
 739     case TOK_SHR:
 740         opc = 5;
 741         goto gen_shift;
 742     case TOK_SAR:
 743         opc = 7;
 744     gen_shift:
 745         opc = 0xc0 | (opc << 3);
 746         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 747             /* constant case */
 748             vswap();
 749             r = gv(RC_INT);
 750             vswap();
 751             c = vtop->c.i & 0x1f;
 752             o(0xc1); /* shl/shr/sar $xxx, r */
 753             o(opc | r);
 754             g(c);
 755         } else {
 756             /* we generate the shift in ecx */
 757             gv2(RC_INT, RC_ECX);
 758             r = vtop[-1].r;
 759             o(0xd3); /* shl/shr/sar %cl, r */
 760             o(opc | r);
 761         }
 762         vtop--;
 763         break;
 764     case '/':
 765     case TOK_UDIV:
 766     case TOK_PDIV:
 767     case '%':
 768     case TOK_UMOD:
 769     case TOK_UMULL:
 770         /* first operand must be in eax */
 771         /* XXX: need better constraint for second operand */
 772         gv2(RC_EAX, RC_ECX);
 773         r = vtop[-1].r;
 774         fr = vtop[0].r;
 775         vtop--;
 776         save_reg(TREG_EDX);
 777         if (op == TOK_UMULL) {
 778             o(0xf7); /* mul fr */
 779             o(0xe0 + fr);
 780             vtop->r2 = TREG_EDX;
 781             r = TREG_EAX;
 782         } else {
 783             if (op == TOK_UDIV || op == TOK_UMOD) {
 784                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 785                 o(0xf0 + fr);
 786             } else {
 787                 o(0xf799); /* cltd, idiv fr, %eax */
 788                 o(0xf8 + fr);
 789             }
 790             if (op == '%' || op == TOK_UMOD)
 791                 r = TREG_EDX;
 792             else
 793                 r = TREG_EAX;
 794         }
 795         vtop->r = r;
 796         break;
 797     default:
 798         opc = 7;
 799         goto gen_op8;
 800     }
 801 }
 802
 803 /* generate a floating point operation 'v = t1 op t2' instruction. The
 804    two operands are guaranted to have the same floating point type */
 805 /* XXX: need to use ST1 too */
 806 void gen_opf(int op)
 807 {
 808     int a, ft, fc, swapped, r;
 809
 810     /* convert constants to memory references */
 811     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 812         vswap();
 813         gv(RC_FLOAT);
 814         vswap();
 815     }
 816     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 817         gv(RC_FLOAT);
 818
 819     /* must put at least one value in the floating point register */
 820     if ((vtop[-1].r & VT_LVAL) &&
 821         (vtop[0].r & VT_LVAL)) {
 822         vswap();
 823         gv(RC_FLOAT);
 824         vswap();
 825     }
 826     swapped = 0;
 827     /* swap the stack if needed so that t1 is the register and t2 is
 828        the memory reference */
 829     if (vtop[-1].r & VT_LVAL) {
 830         vswap();
 831         swapped = 1;
 832     }
 833     if (op >= TOK_ULT && op <= TOK_GT) {
 834         /* load on stack second operand */
 835         load(TREG_ST0, vtop);
 836         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 837         if (op == TOK_GE || op == TOK_GT)
 838             swapped = !swapped;
 839         else if (op == TOK_EQ || op == TOK_NE)
 840             swapped = 0;
 841         if (swapped)
 842             o(0xc9d9); /* fxch %st(1) */
 843         o(0xe9da); /* fucompp */
 844         o(0xe0df); /* fnstsw %ax */
 845         if (op == TOK_EQ) {
 846             o(0x45e480); /* and $0x45, %ah */
 847             o(0x40fC80); /* cmp $0x40, %ah */
 848         } else if (op == TOK_NE) {
 849             o(0x45e480); /* and $0x45, %ah */
 850             o(0x40f480); /* xor $0x40, %ah */
 851             op = TOK_NE;
 852         } else if (op == TOK_GE || op == TOK_LE) {
 853             o(0x05c4f6); /* test $0x05, %ah */
 854             op = TOK_EQ;
 855         } else {
 856             o(0x45c4f6); /* test $0x45, %ah */
 857             op = TOK_EQ;
 858         }
 859         vtop--;
 860         vtop->r = VT_CMP;
 861         vtop->c.i = op;
 862     } else {
 863         /* no memory reference possible for long double operations */
 864         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 865             load(TREG_ST0, vtop);
 866             swapped = !swapped;
 867         }
 868
 869         switch(op) {
 870         default:
 871         case '+':
 872             a = 0;
 873             break;
 874         case '-':
 875             a = 4;
 876             if (swapped)
 877                 a++;
 878             break;
 879         case '*':
 880             a = 1;
 881             break;
 882         case '/':
 883             a = 6;
 884             if (swapped)
 885                 a++;
 886             break;
 887         }
 888         ft = vtop->type.t;
 889         fc = vtop->c.ul;
 890         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 891             o(0xde); /* fxxxp %st, %st(1) */
 892             o(0xc1 + (a << 3));
 893         } else {
 894             /* if saved lvalue, then we must reload it */
 895             r = vtop->r;
 896             if ((r & VT_VALMASK) == VT_LLOCAL) {
 897                 SValue v1;
 898                 r = get_reg(RC_INT);
 899                 v1.type.t = VT_INT;
 900                 v1.r = VT_LOCAL | VT_LVAL;
 901                 v1.c.ul = fc;
 902                 load(r, &v1);
 903                 fc = 0;
 904             }
 905
 906             if ((ft & VT_BTYPE) == VT_DOUBLE)
 907                 o(0xdc);
 908             else
 909                 o(0xd8);
 910             gen_modrm(a, r, vtop->sym, fc);
 911         }
 912         vtop--;
 913     }
 914 }
 915
 916 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 917    and 'long long' cases. */
 918 void gen_cvt_itof(int t)
 919 {
 920     save_reg(TREG_ST0);
 921     gv(RC_INT);
 922     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 923         /* signed long long to float/double/long double (unsigned case
 924            is handled generically) */
 925         o(0x50 + vtop->r2); /* push r2 */
 926         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 927         o(0x242cdf); /* fildll (%esp) */
 928         o(0x08c483); /* add $8, %esp */
 929     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 930                (VT_INT | VT_UNSIGNED)) {
 931         /* unsigned int to float/double/long double */
 932         o(0x6a); /* push $0 */
 933         g(0x00);
 934         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 935         o(0x242cdf); /* fildll (%esp) */
 936         o(0x08c483); /* add $8, %esp */
 937     } else {
 938         /* int to float/double/long double */
 939         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 940         o(0x2404db); /* fildl (%esp) */
 941         o(0x04c483); /* add $4, %esp */
 942     }
 943     vtop->r = TREG_ST0;
 944 }
 945
 946 /* convert fp to int 't' type */
 947 /* XXX: handle long long case */
 948 void gen_cvt_ftoi(int t)
 949 {
 950     int r, r2, size;
 951     Sym *sym;
 952     CType ushort_type;
 953
 954     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 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 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 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 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 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