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     SValue v2;
 219     sv = pe_getimport(sv, &v2);
 220 #endif
 221
 222     fr = sv->r;
 223     ft = sv->type.t;
 224     fc = sv->c.ul;
 225
 226     v = fr & VT_VALMASK;
 227     if (fr & VT_LVAL) {
 228         if (v == VT_LLOCAL) {
 229             v1.type.t = VT_INT;
 230             v1.r = VT_LOCAL | VT_LVAL;
 231             v1.c.ul = fc;
 232             fr = r;
 233             if (!(reg_classes[fr] & RC_INT))
 234                 fr = get_reg(RC_INT);
 235             load(fr, &v1);
 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             if (fc) {
 264                 o(0x8d); /* lea xxx(%ebp), r */
 265                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 266             } else {
 267                 o(0x89);
 268                 o(0xe8 + r); /* mov %ebp, r */
 269             }
 270         } else if (v == VT_CMP) {
 271             oad(0xb8 + r, 0); /* mov $0, r */
 272             o(0x0f); /* setxx %br */
 273             o(fc);
 274             o(0xc0 + r);
 275         } else if (v == VT_JMP || v == VT_JMPI) {
 276             t = v & 1;
 277             oad(0xb8 + r, t); /* mov $1, r */
 278             o(0x05eb); /* jmp after */
 279             gsym(fc);
 280             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 281         } else if (v != r) {
 282             o(0x89);
 283             o(0xc0 + r + v * 8); /* mov v, r */
 284         }
 285     }
 286 }
 287
 288 /* store register 'r' in lvalue 'v' */
 289 ST_FUNC void store(int r, SValue *v)
 290 {
 291     int fr, bt, ft, fc;
 292
 293 #ifdef TCC_TARGET_PE
 294     SValue v2;
 295     v = pe_getimport(v, &v2);
 296 #endif
 297
 298     ft = v->type.t;
 299     fc = v->c.ul;
 300     fr = v->r & VT_VALMASK;
 301     bt = ft & VT_BTYPE;
 302     /* XXX: incorrect if float reg to reg */
 303     if (bt == VT_FLOAT) {
 304         o(0xd9); /* fsts */
 305         r = 2;
 306     } else if (bt == VT_DOUBLE) {
 307         o(0xdd); /* fstpl */
 308         r = 2;
 309     } else if (bt == VT_LDOUBLE) {
 310         o(0xc0d9); /* fld %st(0) */
 311         o(0xdb); /* fstpt */
 312         r = 7;
 313     } else {
 314         if (bt == VT_SHORT)
 315             o(0x66);
 316         if (bt == VT_BYTE || bt == VT_BOOL)
 317             o(0x88);
 318         else
 319             o(0x89);
 320     }
 321     if (fr == VT_CONST ||
 322         fr == VT_LOCAL ||
 323         (v->r & VT_LVAL)) {
 324         gen_modrm(r, v->r, v->sym, fc);
 325     } else if (fr != r) {
 326         o(0xc0 + fr + r * 8); /* mov r, fr */
 327     }
 328 }
 329
 330 static void gadd_sp(int val)
 331 {
 332     if (val == (char)val) {
 333         o(0xc483);
 334         g(val);
 335     } else {
 336         oad(0xc481, val); /* add $xxx, %esp */
 337     }
 338 }
 339
 340 /* 'is_jmp' is '1' if it is a jump */
 341 static void gcall_or_jmp(int is_jmp)
 342 {
 343     int r;
 344     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 345         /* constant case */
 346         if (vtop->r & VT_SYM) {
 347             /* relocation case */
 348             greloc(cur_text_section, vtop->sym,
 349                    ind + 1, R_386_PC32);
 350         } else {
 351             /* put an empty PC32 relocation */
 352             put_elf_reloc(symtab_section, cur_text_section,
 353                           ind + 1, R_386_PC32, 0);
 354         }
 355         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 356     } else {
 357         /* otherwise, indirect call */
 358         r = gv(RC_INT);
 359         o(0xff); /* call/jmp *r */
 360         o(0xd0 + r + (is_jmp << 4));
 361     }
 362 }
 363
 364 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 365 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 366
 367 /* Generate function call. The function address is pushed first, then
 368    all the parameters in call order. This functions pops all the
 369    parameters and the function address. */
 370 ST_FUNC void gfunc_call(int nb_args)
 371 {
 372     int size, align, r, args_size, i, func_call;
 373     Sym *func_sym;
 374
 375     args_size = 0;
 376     for(i = 0;i < nb_args; i++) {
 377         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 378             size = type_size(&vtop->type, &align);
 379             /* align to stack align size */
 380             size = (size + 3) & ~3;
 381             /* allocate the necessary size on stack */
 382             oad(0xec81, size); /* sub $xxx, %esp */
 383             /* generate structure store */
 384             r = get_reg(RC_INT);
 385             o(0x89); /* mov %esp, r */
 386             o(0xe0 + r);
 387             vset(&vtop->type, r | VT_LVAL, 0);
 388             vswap();
 389             vstore();
 390             args_size += size;
 391         } else if (is_float(vtop->type.t)) {
 392             gv(RC_FLOAT); /* only one float register */
 393             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 394                 size = 4;
 395             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 396                 size = 8;
 397             else
 398                 size = 12;
 399             oad(0xec81, size); /* sub $xxx, %esp */
 400             if (size == 12)
 401                 o(0x7cdb);
 402             else
 403                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 404             g(0x24);
 405             g(0x00);
 406             args_size += size;
 407         } else {
 408             /* simple type (currently always same size) */
 409             /* XXX: implicit cast ? */
 410             r = gv(RC_INT);
 411             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 412                 size = 8;
 413                 o(0x50 + vtop->r2); /* push r */
 414             } else {
 415                 size = 4;
 416             }
 417             o(0x50 + r); /* push r */
 418             args_size += size;
 419         }
 420         vtop--;
 421     }
 422     save_regs(0); /* save used temporary registers */
 423     func_sym = vtop->type.ref;
 424     func_call = FUNC_CALL(func_sym->r);
 425     /* fast call case */
 426     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 427         func_call == FUNC_FASTCALLW) {
 428         int fastcall_nb_regs;
 429         uint8_t *fastcall_regs_ptr;
 430         if (func_call == FUNC_FASTCALLW) {
 431             fastcall_regs_ptr = fastcallw_regs;
 432             fastcall_nb_regs = 2;
 433         } else {
 434             fastcall_regs_ptr = fastcall_regs;
 435             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 436         }
 437         for(i = 0;i < fastcall_nb_regs; i++) {
 438             if (args_size <= 0)
 439                 break;
 440             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 441             /* XXX: incorrect for struct/floats */
 442             args_size -= 4;
 443         }
 444     }
 445     gcall_or_jmp(0);
 446
 447 #ifdef TCC_TARGET_PE
 448     if ((func_sym->type.t & VT_BTYPE) == VT_STRUCT)
 449         args_size -= 4;
 450 #endif
 451     if (args_size && func_call != FUNC_STDCALL)
 452         gadd_sp(args_size);
 453     vtop--;
 454 }
 455
 456 #ifdef TCC_TARGET_PE
 457 #define FUNC_PROLOG_SIZE 10
 458 #else
 459 #define FUNC_PROLOG_SIZE 9
 460 #endif
 461
 462 /* generate function prolog of type 't' */
 463 ST_FUNC void gfunc_prolog(CType *func_type)
 464 {
 465     int addr, align, size, func_call, fastcall_nb_regs;
 466     int param_index, param_addr;
 467     uint8_t *fastcall_regs_ptr;
 468     Sym *sym;
 469     CType *type;
 470
 471     sym = func_type->ref;
 472     func_call = FUNC_CALL(sym->r);
 473     addr = 8;
 474     loc = 0;
 475     func_vc = 0;
 476
 477     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 478         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 479         fastcall_regs_ptr = fastcall_regs;
 480     } else if (func_call == FUNC_FASTCALLW) {
 481         fastcall_nb_regs = 2;
 482         fastcall_regs_ptr = fastcallw_regs;
 483     } else {
 484         fastcall_nb_regs = 0;
 485         fastcall_regs_ptr = NULL;
 486     }
 487     param_index = 0;
 488
 489     ind += FUNC_PROLOG_SIZE;
 490     func_sub_sp_offset = ind;
 491     /* if the function returns a structure, then add an
 492        implicit pointer parameter */
 493     func_vt = sym->type;
 494     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 495         /* XXX: fastcall case ? */
 496         func_vc = addr;
 497         addr += 4;
 498         param_index++;
 499     }
 500     /* define parameters */
 501     while ((sym = sym->next) != NULL) {
 502         type = &sym->type;
 503         size = type_size(type, &align);
 504         size = (size + 3) & ~3;
 505 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 506         /* structs are passed as pointer */
 507         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 508             size = 4;
 509         }
 510 #endif
 511         if (param_index < fastcall_nb_regs) {
 512             /* save FASTCALL register */
 513             loc -= 4;
 514             o(0x89);     /* movl */
 515             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 516             param_addr = loc;
 517         } else {
 518             param_addr = addr;
 519             addr += size;
 520         }
 521         sym_push(sym->v & ~SYM_FIELD, type,
 522                  VT_LOCAL | lvalue_type(type->t), param_addr);
 523         param_index++;
 524     }
 525     func_ret_sub = 0;
 526     /* pascal type call ? */
 527     if (func_call == FUNC_STDCALL)
 528         func_ret_sub = addr - 8;
 529 #ifdef TCC_TARGET_PE
 530     else if (func_vc)
 531         func_ret_sub = 4;
 532 #endif
 533
 534 #ifdef CONFIG_TCC_BCHECK
 535     /* leave some room for bound checking code */
 536     if (tcc_state->do_bounds_check) {
 537         oad(0xb8, 0); /* lbound section pointer */
 538         oad(0xb8, 0); /* call to function */
 539         func_bound_offset = lbounds_section->data_offset;
 540     }
 541 #endif
 542 }
 543
 544 /* generate function epilog */
 545 ST_FUNC void gfunc_epilog(void)
 546 {
 547     int v, saved_ind;
 548
 549 #ifdef CONFIG_TCC_BCHECK
 550     if (tcc_state->do_bounds_check
 551      && func_bound_offset != lbounds_section->data_offset) {
 552         int saved_ind;
 553         int *bounds_ptr;
 554         Sym *sym, *sym_data;
 555         /* add end of table info */
 556         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 557         *bounds_ptr = 0;
 558         /* generate bound local allocation */
 559         saved_ind = ind;
 560         ind = func_sub_sp_offset;
 561         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 562                                func_bound_offset, lbounds_section->data_offset);
 563         greloc(cur_text_section, sym_data,
 564                ind + 1, R_386_32);
 565         oad(0xb8, 0); /* mov %eax, xxx */
 566         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 567         greloc(cur_text_section, sym,
 568                ind + 1, R_386_PC32);
 569         oad(0xe8, -4);
 570         ind = saved_ind;
 571         /* generate bound check local freeing */
 572         o(0x5250); /* save returned value, if any */
 573         greloc(cur_text_section, sym_data,
 574                ind + 1, R_386_32);
 575         oad(0xb8, 0); /* mov %eax, xxx */
 576         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 577         greloc(cur_text_section, sym,
 578                ind + 1, R_386_PC32);
 579         oad(0xe8, -4);
 580         o(0x585a); /* restore returned value, if any */
 581     }
 582 #endif
 583     o(0xc9); /* leave */
 584     if (func_ret_sub == 0) {
 585         o(0xc3); /* ret */
 586     } else {
 587         o(0xc2); /* ret n */
 588         g(func_ret_sub);
 589         g(func_ret_sub >> 8);
 590     }
 591     /* align local size to word & save local variables */
 592
 593     v = (-loc + 3) & -4;
 594     saved_ind = ind;
 595     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 596 #ifdef TCC_TARGET_PE
 597     if (v >= 4096) {
 598         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 599         oad(0xb8, v); /* mov stacksize, %eax */
 600         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 601         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 602     } else
 603 #endif
 604     {
 605         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 606         o(0xec81);  /* sub esp, stacksize */
 607         gen_le32(v);
 608 #if FUNC_PROLOG_SIZE == 10
 609         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 610 #endif
 611     }
 612     ind = saved_ind;
 613 }
 614
 615 /* generate a jump to a label */
 616 ST_FUNC int gjmp(int t)
 617 {
 618     return psym(0xe9, t);
 619 }
 620
 621 /* generate a jump to a fixed address */
 622 ST_FUNC void gjmp_addr(int a)
 623 {
 624     int r;
 625     r = a - ind - 2;
 626     if (r == (char)r) {
 627         g(0xeb);
 628         g(r);
 629     } else {
 630         oad(0xe9, a - ind - 5);
 631     }
 632 }
 633
 634 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 635 ST_FUNC int gtst(int inv, int t)
 636 {
 637     int v, *p;
 638
 639     v = vtop->r & VT_VALMASK;
 640     if (v == VT_CMP) {
 641         /* fast case : can jump directly since flags are set */
 642         g(0x0f);
 643         t = psym((vtop->c.i - 16) ^ inv, t);
 644     } else if (v == VT_JMP || v == VT_JMPI) {
 645         /* && or || optimization */
 646         if ((v & 1) == inv) {
 647             /* insert vtop->c jump list in t */
 648             p = &vtop->c.i;
 649             while (*p != 0)
 650                 p = (int *)(cur_text_section->data + *p);
 651             *p = t;
 652             t = vtop->c.i;
 653         } else {
 654             t = gjmp(t);
 655             gsym(vtop->c.i);
 656         }
 657     } else {
 658         if (is_float(vtop->type.t) ||
 659             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 660             vpushi(0);
 661             gen_op(TOK_NE);
 662         }
 663         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 664             /* constant jmp optimization */
 665             if ((vtop->c.i != 0) != inv)
 666                 t = gjmp(t);
 667         } else {
 668             v = gv(RC_INT);
 669             o(0x85);
 670             o(0xc0 + v * 9);
 671             g(0x0f);
 672             t = psym(0x85 ^ inv, t);
 673         }
 674     }
 675     vtop--;
 676     return t;
 677 }
 678
 679 /* generate an integer binary operation */
 680 ST_FUNC void gen_opi(int op)
 681 {
 682     int r, fr, opc, c;
 683
 684     switch(op) {
 685     case '+':
 686     case TOK_ADDC1: /* add with carry generation */
 687         opc = 0;
 688     gen_op8:
 689         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 690             /* constant case */
 691             vswap();
 692             r = gv(RC_INT);
 693             vswap();
 694             c = vtop->c.i;
 695             if (c == (char)c) {
 696                 /* generate inc and dec for smaller code */
 697                 if (c==1 && opc==0) {
 698                     o (0x40 | r); // inc
 699                 } else if (c==1 && opc==5) {
 700                     o (0x48 | r); // dec
 701                 } else {
 702                     o(0x83);
 703                     o(0xc0 | (opc << 3) | r);
 704                     g(c);
 705                 }
 706             } else {
 707                 o(0x81);
 708                 oad(0xc0 | (opc << 3) | r, c);
 709             }
 710         } else {
 711             gv2(RC_INT, RC_INT);
 712             r = vtop[-1].r;
 713             fr = vtop[0].r;
 714             o((opc << 3) | 0x01);
 715             o(0xc0 + r + fr * 8);
 716         }
 717         vtop--;
 718         if (op >= TOK_ULT && op <= TOK_GT) {
 719             vtop->r = VT_CMP;
 720             vtop->c.i = op;
 721         }
 722         break;
 723     case '-':
 724     case TOK_SUBC1: /* sub with carry generation */
 725         opc = 5;
 726         goto gen_op8;
 727     case TOK_ADDC2: /* add with carry use */
 728         opc = 2;
 729         goto gen_op8;
 730     case TOK_SUBC2: /* sub with carry use */
 731         opc = 3;
 732         goto gen_op8;
 733     case '&':
 734         opc = 4;
 735         goto gen_op8;
 736     case '^':
 737         opc = 6;
 738         goto gen_op8;
 739     case '|':
 740         opc = 1;
 741         goto gen_op8;
 742     case '*':
 743         gv2(RC_INT, RC_INT);
 744         r = vtop[-1].r;
 745         fr = vtop[0].r;
 746         vtop--;
 747         o(0xaf0f); /* imul fr, r */
 748         o(0xc0 + fr + r * 8);
 749         break;
 750     case TOK_SHL:
 751         opc = 4;
 752         goto gen_shift;
 753     case TOK_SHR:
 754         opc = 5;
 755         goto gen_shift;
 756     case TOK_SAR:
 757         opc = 7;
 758     gen_shift:
 759         opc = 0xc0 | (opc << 3);
 760         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 761             /* constant case */
 762             vswap();
 763             r = gv(RC_INT);
 764             vswap();
 765             c = vtop->c.i & 0x1f;
 766             o(0xc1); /* shl/shr/sar $xxx, r */
 767             o(opc | r);
 768             g(c);
 769         } else {
 770             /* we generate the shift in ecx */
 771             gv2(RC_INT, RC_ECX);
 772             r = vtop[-1].r;
 773             o(0xd3); /* shl/shr/sar %cl, r */
 774             o(opc | r);
 775         }
 776         vtop--;
 777         break;
 778     case '/':
 779     case TOK_UDIV:
 780     case TOK_PDIV:
 781     case '%':
 782     case TOK_UMOD:
 783     case TOK_UMULL:
 784         /* first operand must be in eax */
 785         /* XXX: need better constraint for second operand */
 786         gv2(RC_EAX, RC_ECX);
 787         r = vtop[-1].r;
 788         fr = vtop[0].r;
 789         vtop--;
 790         save_reg(TREG_EDX);
 791         if (op == TOK_UMULL) {
 792             o(0xf7); /* mul fr */
 793             o(0xe0 + fr);
 794             vtop->r2 = TREG_EDX;
 795             r = TREG_EAX;
 796         } else {
 797             if (op == TOK_UDIV || op == TOK_UMOD) {
 798                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 799                 o(0xf0 + fr);
 800             } else {
 801                 o(0xf799); /* cltd, idiv fr, %eax */
 802                 o(0xf8 + fr);
 803             }
 804             if (op == '%' || op == TOK_UMOD)
 805                 r = TREG_EDX;
 806             else
 807                 r = TREG_EAX;
 808         }
 809         vtop->r = r;
 810         break;
 811     default:
 812         opc = 7;
 813         goto gen_op8;
 814     }
 815 }
 816
 817 /* generate a floating point operation 'v = t1 op t2' instruction. The
 818    two operands are guaranted to have the same floating point type */
 819 /* XXX: need to use ST1 too */
 820 ST_FUNC void gen_opf(int op)
 821 {
 822     int a, ft, fc, swapped, r;
 823
 824     /* convert constants to memory references */
 825     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 826         vswap();
 827         gv(RC_FLOAT);
 828         vswap();
 829     }
 830     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 831         gv(RC_FLOAT);
 832
 833     /* must put at least one value in the floating point register */
 834     if ((vtop[-1].r & VT_LVAL) &&
 835         (vtop[0].r & VT_LVAL)) {
 836         vswap();
 837         gv(RC_FLOAT);
 838         vswap();
 839     }
 840     swapped = 0;
 841     /* swap the stack if needed so that t1 is the register and t2 is
 842        the memory reference */
 843     if (vtop[-1].r & VT_LVAL) {
 844         vswap();
 845         swapped = 1;
 846     }
 847     if (op >= TOK_ULT && op <= TOK_GT) {
 848         /* load on stack second operand */
 849         load(TREG_ST0, vtop);
 850         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 851         if (op == TOK_GE || op == TOK_GT)
 852             swapped = !swapped;
 853         else if (op == TOK_EQ || op == TOK_NE)
 854             swapped = 0;
 855         if (swapped)
 856             o(0xc9d9); /* fxch %st(1) */
 857         o(0xe9da); /* fucompp */
 858         o(0xe0df); /* fnstsw %ax */
 859         if (op == TOK_EQ) {
 860             o(0x45e480); /* and $0x45, %ah */
 861             o(0x40fC80); /* cmp $0x40, %ah */
 862         } else if (op == TOK_NE) {
 863             o(0x45e480); /* and $0x45, %ah */
 864             o(0x40f480); /* xor $0x40, %ah */
 865             op = TOK_NE;
 866         } else if (op == TOK_GE || op == TOK_LE) {
 867             o(0x05c4f6); /* test $0x05, %ah */
 868             op = TOK_EQ;
 869         } else {
 870             o(0x45c4f6); /* test $0x45, %ah */
 871             op = TOK_EQ;
 872         }
 873         vtop--;
 874         vtop->r = VT_CMP;
 875         vtop->c.i = op;
 876     } else {
 877         /* no memory reference possible for long double operations */
 878         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 879             load(TREG_ST0, vtop);
 880             swapped = !swapped;
 881         }
 882
 883         switch(op) {
 884         default:
 885         case '+':
 886             a = 0;
 887             break;
 888         case '-':
 889             a = 4;
 890             if (swapped)
 891                 a++;
 892             break;
 893         case '*':
 894             a = 1;
 895             break;
 896         case '/':
 897             a = 6;
 898             if (swapped)
 899                 a++;
 900             break;
 901         }
 902         ft = vtop->type.t;
 903         fc = vtop->c.ul;
 904         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 905             o(0xde); /* fxxxp %st, %st(1) */
 906             o(0xc1 + (a << 3));
 907         } else {
 908             /* if saved lvalue, then we must reload it */
 909             r = vtop->r;
 910             if ((r & VT_VALMASK) == VT_LLOCAL) {
 911                 SValue v1;
 912                 r = get_reg(RC_INT);
 913                 v1.type.t = VT_INT;
 914                 v1.r = VT_LOCAL | VT_LVAL;
 915                 v1.c.ul = fc;
 916                 load(r, &v1);
 917                 fc = 0;
 918             }
 919
 920             if ((ft & VT_BTYPE) == VT_DOUBLE)
 921                 o(0xdc);
 922             else
 923                 o(0xd8);
 924             gen_modrm(a, r, vtop->sym, fc);
 925         }
 926         vtop--;
 927     }
 928 }
 929
 930 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 931    and 'long long' cases. */
 932 ST_FUNC void gen_cvt_itof(int t)
 933 {
 934     save_reg(TREG_ST0);
 935     gv(RC_INT);
 936     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 937         /* signed long long to float/double/long double (unsigned case
 938            is handled generically) */
 939         o(0x50 + vtop->r2); /* push r2 */
 940         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 941         o(0x242cdf); /* fildll (%esp) */
 942         o(0x08c483); /* add $8, %esp */
 943     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 944                (VT_INT | VT_UNSIGNED)) {
 945         /* unsigned int to float/double/long double */
 946         o(0x6a); /* push $0 */
 947         g(0x00);
 948         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 949         o(0x242cdf); /* fildll (%esp) */
 950         o(0x08c483); /* add $8, %esp */
 951     } else {
 952         /* int to float/double/long double */
 953         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 954         o(0x2404db); /* fildl (%esp) */
 955         o(0x04c483); /* add $4, %esp */
 956     }
 957     vtop->r = TREG_ST0;
 958 }
 959
 960 /* convert fp to int 't' type */
 961 /* XXX: handle long long case */
 962 ST_FUNC void gen_cvt_ftoi(int t)
 963 {
 964     int r, r2, size;
 965     Sym *sym;
 966     CType ushort_type;
 967
 968     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 969     ushort_type.ref = 0;
 970
 971     gv(RC_FLOAT);
 972     if (t != VT_INT)
 973         size = 8;
 974     else
 975         size = 4;
 976
 977     o(0x2dd9); /* ldcw xxx */
 978     sym = external_global_sym(TOK___tcc_int_fpu_control,
 979                               &ushort_type, VT_LVAL);
 980     greloc(cur_text_section, sym,
 981            ind, R_386_32);
 982     gen_le32(0);
 983
 984     oad(0xec81, size); /* sub $xxx, %esp */
 985     if (size == 4)
 986         o(0x1cdb); /* fistpl */
 987     else
 988         o(0x3cdf); /* fistpll */
 989     o(0x24);
 990     o(0x2dd9); /* ldcw xxx */
 991     sym = external_global_sym(TOK___tcc_fpu_control,
 992                               &ushort_type, VT_LVAL);
 993     greloc(cur_text_section, sym,
 994            ind, R_386_32);
 995     gen_le32(0);
 996
 997     r = get_reg(RC_INT);
 998     o(0x58 + r); /* pop r */
 999     if (size == 8) {
1000         if (t == VT_LLONG) {
1001             vtop->r = r; /* mark reg as used */
1002             r2 = get_reg(RC_INT);
1003             o(0x58 + r2); /* pop r2 */
1004             vtop->r2 = r2;
1005         } else {
1006             o(0x04c483); /* add $4, %esp */
1007         }
1008     }
1009     vtop->r = r;
1010 }
1011
1012 /* convert from one floating point type to another */
1013 ST_FUNC void gen_cvt_ftof(int t)
1014 {
1015     /* all we have to do on i386 is to put the float in a register */
1016     gv(RC_FLOAT);
1017 }
1018
1019 /* computed goto support */
1020 ST_FUNC void ggoto(void)
1021 {
1022     gcall_or_jmp(1);
1023     vtop--;
1024 }
1025
1026 /* bound check support functions */
1027 #ifdef CONFIG_TCC_BCHECK
1028
1029 /* generate a bounded pointer addition */
1030 ST_FUNC void gen_bounded_ptr_add(void)
1031 {
1032     Sym *sym;
1033
1034     /* prepare fast i386 function call (args in eax and edx) */
1035     gv2(RC_EAX, RC_EDX);
1036     /* save all temporary registers */
1037     vtop -= 2;
1038     save_regs(0);
1039     /* do a fast function call */
1040     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
1041     greloc(cur_text_section, sym,
1042            ind + 1, R_386_PC32);
1043     oad(0xe8, -4);
1044     /* returned pointer is in eax */
1045     vtop++;
1046     vtop->r = TREG_EAX | VT_BOUNDED;
1047     /* address of bounding function call point */
1048     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1049 }
1050
1051 /* patch pointer addition in vtop so that pointer dereferencing is
1052    also tested */
1053 ST_FUNC void gen_bounded_ptr_deref(void)
1054 {
1055     int func;
1056     int size, align;
1057     Elf32_Rel *rel;
1058     Sym *sym;
1059
1060     size = 0;
1061     /* XXX: put that code in generic part of tcc */
1062     if (!is_float(vtop->type.t)) {
1063         if (vtop->r & VT_LVAL_BYTE)
1064             size = 1;
1065         else if (vtop->r & VT_LVAL_SHORT)
1066             size = 2;
1067     }
1068     if (!size)
1069         size = type_size(&vtop->type, &align);
1070     switch(size) {
1071     case  1: func = TOK___bound_ptr_indir1; break;
1072     case  2: func = TOK___bound_ptr_indir2; break;
1073     case  4: func = TOK___bound_ptr_indir4; break;
1074     case  8: func = TOK___bound_ptr_indir8; break;
1075     case 12: func = TOK___bound_ptr_indir12; break;
1076     case 16: func = TOK___bound_ptr_indir16; break;
1077     default:
1078         tcc_error("unhandled size when dereferencing bounded pointer");
1079         func = 0;
1080         break;
1081     }
1082
1083     /* patch relocation */
1084     /* XXX: find a better solution ? */
1085     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1086     sym = external_global_sym(func, &func_old_type, 0);
1087     if (!sym->c)
1088         put_extern_sym(sym, NULL, 0, 0);
1089     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1090 }
1091 #endif
1092
1093 /* end of X86 code generator */
1094 /*************************************************************/
1095 #endif
1096 /*************************************************************/