i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001-2004 Fabrice Bellard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  19  */
  20
  21 /* number of available registers */
  22 #define NB_REGS             4
  23
  24 /* a register can belong to several classes. The classes must be
  25    sorted from more general to more precise (see gv2() code which does
  26    assumptions on it). */
  27 #define RC_INT     0x0001 /* generic integer register */
  28 #define RC_FLOAT   0x0002 /* generic float register */
  29 #define RC_EAX     0x0004
  30 #define RC_ST0     0x0008
  31 #define RC_ECX     0x0010
  32 #define RC_EDX     0x0020
  33 #define RC_IRET    RC_EAX /* function return: integer register */
  34 #define RC_LRET    RC_EDX /* function return: second integer register */
  35 #define RC_FRET    RC_ST0 /* function return: float register */
  36
  37 /* pretty names for the registers */
  38 enum {
  39     TREG_EAX = 0,
  40     TREG_ECX,
  41     TREG_EDX,
  42     TREG_ST0,
  43 };
  44
  45 int reg_classes[NB_REGS] = {
  46     /* eax */ RC_INT | RC_EAX,
  47     /* ecx */ RC_INT | RC_ECX,
  48     /* edx */ RC_INT | RC_EDX,
  49     /* st0 */ RC_FLOAT | RC_ST0,
  50 };
  51
  52 /* return registers for function */
  53 #define REG_IRET TREG_EAX /* single word int return register */
  54 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  55 #define REG_FRET TREG_ST0 /* float return register */
  56
  57 /* defined if function parameters must be evaluated in reverse order */
  58 #define INVERT_FUNC_PARAMS
  59
  60 /* defined if structures are passed as pointers. Otherwise structures
  61    are directly pushed on stack. */
  62 //#define FUNC_STRUCT_PARAM_AS_PTR
  63
  64 /* pointer size, in bytes */
  65 #define PTR_SIZE 4
  66
  67 /* long double size and alignment, in bytes */
  68 #define LDOUBLE_SIZE  12
  69 #define LDOUBLE_ALIGN 4
  70 /* maximum alignment (for aligned attribute support) */
  71 #define MAX_ALIGN     8
  72
  73 /******************************************************/
  74 /* ELF defines */
  75
  76 #define EM_TCC_TARGET EM_386
  77
  78 /* relocation type for 32 bit data relocation */
  79 #define R_DATA_32   R_386_32
  80 #define R_JMP_SLOT  R_386_JMP_SLOT
  81 #define R_COPY      R_386_COPY
  82
  83 #define ELF_START_ADDR 0x08048000
  84 #define ELF_PAGE_SIZE  0x1000
  85
  86 /******************************************************/
  87
  88 static unsigned long func_sub_sp_offset;
  89 static unsigned long func_bound_offset;
  90 static int func_ret_sub;
  91
  92 /* XXX: make it faster ? */
  93 void g(int c)
  94 {
  95     int ind1;
  96     ind1 = ind + 1;
  97     if (ind1 > cur_text_section->data_allocated)
  98         section_realloc(cur_text_section, ind1);
  99     cur_text_section->data[ind] = c;
 100     ind = ind1;
 101 }
 102
 103 void o(unsigned int c)
 104 {
 105     while (c) {
 106         g(c);
 107         c = c >> 8;
 108     }
 109 }
 110
 111 void gen_le32(int c)
 112 {
 113     g(c);
 114     g(c >> 8);
 115     g(c >> 16);
 116     g(c >> 24);
 117 }
 118
 119 /* output a symbol and patch all calls to it */
 120 void gsym_addr(int t, int a)
 121 {
 122     int n, *ptr;
 123     while (t) {
 124         ptr = (int *)(cur_text_section->data + t);
 125         n = *ptr; /* next value */
 126         *ptr = a - t - 4;
 127         t = n;
 128     }
 129 }
 130
 131 void gsym(int t)
 132 {
 133     gsym_addr(t, ind);
 134 }
 135
 136 /* psym is used to put an instruction with a data field which is a
 137    reference to a symbol. It is in fact the same as oad ! */
 138 #define psym oad
 139
 140 /* instruction + 4 bytes data. Return the address of the data */
 141 static int oad(int c, int s)
 142 {
 143     int ind1;
 144
 145     o(c);
 146     ind1 = ind + 4;
 147     if (ind1 > cur_text_section->data_allocated)
 148         section_realloc(cur_text_section, ind1);
 149     *(int *)(cur_text_section->data + ind) = s;
 150     s = ind;
 151     ind = ind1;
 152     return s;
 153 }
 154
 155 /* output constant with relocation if 'r & VT_SYM' is true */
 156 static void gen_addr32(int r, Sym *sym, int c)
 157 {
 158     if (r & VT_SYM)
 159         greloc(cur_text_section, sym, ind, R_386_32);
 160     gen_le32(c);
 161 }
 162
 163 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 164    opcode bits */
 165 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 166 {
 167     op_reg = op_reg << 3;
 168     if ((r & VT_VALMASK) == VT_CONST) {
 169         /* constant memory reference */
 170         o(0x05 | op_reg);
 171         gen_addr32(r, sym, c);
 172     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 173         /* currently, we use only ebp as base */
 174         if (c == (char)c) {
 175             /* short reference */
 176             o(0x45 | op_reg);
 177             g(c);
 178         } else {
 179             oad(0x85 | op_reg, c);
 180         }
 181     } else {
 182         g(0x00 | op_reg | (r & VT_VALMASK));
 183     }
 184 }
 185
 186
 187 /* load 'r' from value 'sv' */
 188 void load(int r, SValue *sv)
 189 {
 190     int v, t, ft, fc, fr;
 191     SValue v1;
 192
 193     fr = sv->r;
 194     ft = sv->type.t;
 195     fc = sv->c.ul;
 196
 197     v = fr & VT_VALMASK;
 198     if (fr & VT_LVAL) {
 199         if (v == VT_LLOCAL) {
 200             v1.type.t = VT_INT;
 201             v1.r = VT_LOCAL | VT_LVAL;
 202             v1.c.ul = fc;
 203             load(r, &v1);
 204             fr = r;
 205         }
 206         if ((ft & VT_BTYPE) == VT_FLOAT) {
 207             o(0xd9); /* flds */
 208             r = 0;
 209         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 210             o(0xdd); /* fldl */
 211             r = 0;
 212         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 213             o(0xdb); /* fldt */
 214             r = 5;
 215         } else if ((ft & VT_TYPE) == VT_BYTE) {
 216             o(0xbe0f);   /* movsbl */
 217         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 218             o(0xb60f);   /* movzbl */
 219         } else if ((ft & VT_TYPE) == VT_SHORT) {
 220             o(0xbf0f);   /* movswl */
 221         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 222             o(0xb70f);   /* movzwl */
 223         } else {
 224             o(0x8b);     /* movl */
 225         }
 226         gen_modrm(r, fr, sv->sym, fc);
 227     } else {
 228         if (v == VT_CONST) {
 229             o(0xb8 + r); /* mov $xx, r */
 230             gen_addr32(fr, sv->sym, fc);
 231         } else if (v == VT_LOCAL) {
 232             o(0x8d); /* lea xxx(%ebp), r */
 233             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 234         } else if (v == VT_CMP) {
 235             oad(0xb8 + r, 0); /* mov $0, r */
 236             o(0x0f); /* setxx %br */
 237             o(fc);
 238             o(0xc0 + r);
 239         } else if (v == VT_JMP || v == VT_JMPI) {
 240             t = v & 1;
 241             oad(0xb8 + r, t); /* mov $1, r */
 242             o(0x05eb); /* jmp after */
 243             gsym(fc);
 244             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 245         } else if (v != r) {
 246             o(0x89);
 247             o(0xc0 + r + v * 8); /* mov v, r */
 248         }
 249     }
 250 }
 251
 252 /* store register 'r' in lvalue 'v' */
 253 void store(int r, SValue *v)
 254 {
 255     int fr, bt, ft, fc;
 256
 257     ft = v->type.t;
 258     fc = v->c.ul;
 259     fr = v->r & VT_VALMASK;
 260     bt = ft & VT_BTYPE;
 261     /* XXX: incorrect if float reg to reg */
 262     if (bt == VT_FLOAT) {
 263         o(0xd9); /* fsts */
 264         r = 2;
 265     } else if (bt == VT_DOUBLE) {
 266         o(0xdd); /* fstpl */
 267         r = 2;
 268     } else if (bt == VT_LDOUBLE) {
 269         o(0xc0d9); /* fld %st(0) */
 270         o(0xdb); /* fstpt */
 271         r = 7;
 272     } else {
 273         if (bt == VT_SHORT)
 274             o(0x66);
 275         if (bt == VT_BYTE || bt == VT_BOOL)
 276             o(0x88);
 277         else
 278             o(0x89);
 279     }
 280     if (fr == VT_CONST ||
 281         fr == VT_LOCAL ||
 282         (v->r & VT_LVAL)) {
 283         gen_modrm(r, v->r, v->sym, fc);
 284     } else if (fr != r) {
 285         o(0xc0 + fr + r * 8); /* mov r, fr */
 286     }
 287 }
 288
 289 static void gadd_sp(int val)
 290 {
 291     if (val == (char)val) {
 292         o(0xc483);
 293         g(val);
 294     } else {
 295         oad(0xc481, val); /* add $xxx, %esp */
 296     }
 297 }
 298
 299 /* 'is_jmp' is '1' if it is a jump */
 300 static void gcall_or_jmp(int is_jmp)
 301 {
 302     int r;
 303     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 304         /* constant case */
 305         if (vtop->r & VT_SYM) {
 306             /* relocation case */
 307             greloc(cur_text_section, vtop->sym,
 308                    ind + 1, R_386_PC32);
 309         } else {
 310             /* put an empty PC32 relocation */
 311             put_elf_reloc(symtab_section, cur_text_section,
 312                           ind + 1, R_386_PC32, 0);
 313         }
 314         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 315     } else {
 316         /* otherwise, indirect call */
 317         r = gv(RC_INT);
 318         o(0xff); /* call/jmp *r */
 319         o(0xd0 + r + (is_jmp << 4));
 320     }
 321 }
 322
 323 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 324 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 325
 326 /* Generate function call. The function address is pushed first, then
 327    all the parameters in call order. This functions pops all the
 328    parameters and the function address. */
 329 void gfunc_call(int nb_args)
 330 {
 331     int size, align, r, args_size, i, func_call;
 332     Sym *func_sym;
 333
 334     args_size = 0;
 335     for(i = 0;i < nb_args; i++) {
 336         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 337             size = type_size(&vtop->type, &align);
 338             /* align to stack align size */
 339             size = (size + 3) & ~3;
 340             /* allocate the necessary size on stack */
 341             oad(0xec81, size); /* sub $xxx, %esp */
 342             /* generate structure store */
 343             r = get_reg(RC_INT);
 344             o(0x89); /* mov %esp, r */
 345             o(0xe0 + r);
 346             vset(&vtop->type, r | VT_LVAL, 0);
 347             vswap();
 348             vstore();
 349             args_size += size;
 350         } else if (is_float(vtop->type.t)) {
 351             gv(RC_FLOAT); /* only one float register */
 352             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 353                 size = 4;
 354             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 355                 size = 8;
 356             else
 357                 size = 12;
 358             oad(0xec81, size); /* sub $xxx, %esp */
 359             if (size == 12)
 360                 o(0x7cdb);
 361             else
 362                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 363             g(0x24);
 364             g(0x00);
 365             args_size += size;
 366         } else {
 367             /* simple type (currently always same size) */
 368             /* XXX: implicit cast ? */
 369             r = gv(RC_INT);
 370             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 371                 size = 8;
 372                 o(0x50 + vtop->r2); /* push r */
 373             } else {
 374                 size = 4;
 375             }
 376             o(0x50 + r); /* push r */
 377             args_size += size;
 378         }
 379         vtop--;
 380     }
 381     save_regs(0); /* save used temporary registers */
 382     func_sym = vtop->type.ref;
 383     func_call = FUNC_CALL(func_sym->r);
 384     /* fast call case */
 385     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 386         func_call == FUNC_FASTCALLW) {
 387         int fastcall_nb_regs;
 388         uint8_t *fastcall_regs_ptr;
 389         if (func_call == FUNC_FASTCALLW) {
 390             fastcall_regs_ptr = fastcallw_regs;
 391             fastcall_nb_regs = 2;
 392         } else {
 393             fastcall_regs_ptr = fastcall_regs;
 394             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 395         }
 396         for(i = 0;i < fastcall_nb_regs; i++) {
 397             if (args_size <= 0)
 398                 break;
 399             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 400             /* XXX: incorrect for struct/floats */
 401             args_size -= 4;
 402         }
 403     }
 404     gcall_or_jmp(0);
 405
 406 #ifdef TCC_TARGET_PE
 407     if ((func_sym->type.t & VT_BTYPE) == VT_STRUCT)
 408         args_size -= 4;
 409 #endif
 410     if (args_size && func_call != FUNC_STDCALL)
 411         gadd_sp(args_size);
 412     vtop--;
 413 }
 414
 415 #ifdef TCC_TARGET_PE
 416 #define FUNC_PROLOG_SIZE 10
 417 #else
 418 #define FUNC_PROLOG_SIZE 9
 419 #endif
 420
 421 /* generate function prolog of type 't' */
 422 void gfunc_prolog(CType *func_type)
 423 {
 424     int addr, align, size, func_call, fastcall_nb_regs;
 425     int param_index, param_addr;
 426     uint8_t *fastcall_regs_ptr;
 427     Sym *sym;
 428     CType *type;
 429
 430     sym = func_type->ref;
 431     func_call = FUNC_CALL(sym->r);
 432     addr = 8;
 433     loc = 0;
 434     func_vc = 0;
 435
 436     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 437         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 438         fastcall_regs_ptr = fastcall_regs;
 439     } else if (func_call == FUNC_FASTCALLW) {
 440         fastcall_nb_regs = 2;
 441         fastcall_regs_ptr = fastcallw_regs;
 442     } else {
 443         fastcall_nb_regs = 0;
 444         fastcall_regs_ptr = NULL;
 445     }
 446     param_index = 0;
 447
 448     ind += FUNC_PROLOG_SIZE;
 449     func_sub_sp_offset = ind;
 450     /* if the function returns a structure, then add an
 451        implicit pointer parameter */
 452     func_vt = sym->type;
 453     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 454         /* XXX: fastcall case ? */
 455         func_vc = addr;
 456         addr += 4;
 457         param_index++;
 458     }
 459     /* define parameters */
 460     while ((sym = sym->next) != NULL) {
 461         type = &sym->type;
 462         size = type_size(type, &align);
 463         size = (size + 3) & ~3;
 464 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 465         /* structs are passed as pointer */
 466         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 467             size = 4;
 468         }
 469 #endif
 470         if (param_index < fastcall_nb_regs) {
 471             /* save FASTCALL register */
 472             loc -= 4;
 473             o(0x89);     /* movl */
 474             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 475             param_addr = loc;
 476         } else {
 477             param_addr = addr;
 478             addr += size;
 479         }
 480         sym_push(sym->v & ~SYM_FIELD, type,
 481                  VT_LOCAL | lvalue_type(type->t), param_addr);
 482         param_index++;
 483     }
 484     func_ret_sub = 0;
 485     /* pascal type call ? */
 486     if (func_call == FUNC_STDCALL)
 487         func_ret_sub = addr - 8;
 488 #ifdef TCC_TARGET_PE
 489     else if (func_vc)
 490         func_ret_sub = 4;
 491 #endif
 492
 493     /* leave some room for bound checking code */
 494     if (tcc_state->do_bounds_check) {
 495         oad(0xb8, 0); /* lbound section pointer */
 496         oad(0xb8, 0); /* call to function */
 497         func_bound_offset = lbounds_section->data_offset;
 498     }
 499 }
 500
 501 /* generate function epilog */
 502 void gfunc_epilog(void)
 503 {
 504     int v, saved_ind;
 505
 506 #ifdef CONFIG_TCC_BCHECK
 507     if (tcc_state->do_bounds_check
 508      && func_bound_offset != lbounds_section->data_offset) {
 509         int saved_ind;
 510         int *bounds_ptr;
 511         Sym *sym, *sym_data;
 512         /* add end of table info */
 513         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 514         *bounds_ptr = 0;
 515         /* generate bound local allocation */
 516         saved_ind = ind;
 517         ind = func_sub_sp_offset;
 518         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 519                                func_bound_offset, lbounds_section->data_offset);
 520         greloc(cur_text_section, sym_data,
 521                ind + 1, R_386_32);
 522         oad(0xb8, 0); /* mov %eax, xxx */
 523         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 524         greloc(cur_text_section, sym,
 525                ind + 1, R_386_PC32);
 526         oad(0xe8, -4);
 527         ind = saved_ind;
 528         /* generate bound check local freeing */
 529         o(0x5250); /* save returned value, if any */
 530         greloc(cur_text_section, sym_data,
 531                ind + 1, R_386_32);
 532         oad(0xb8, 0); /* mov %eax, xxx */
 533         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 534         greloc(cur_text_section, sym,
 535                ind + 1, R_386_PC32);
 536         oad(0xe8, -4);
 537         o(0x585a); /* restore returned value, if any */
 538     }
 539 #endif
 540     o(0xc9); /* leave */
 541     if (func_ret_sub == 0) {
 542         o(0xc3); /* ret */
 543     } else {
 544         o(0xc2); /* ret n */
 545         g(func_ret_sub);
 546         g(func_ret_sub >> 8);
 547     }
 548     /* align local size to word & save local variables */
 549
 550     v = (-loc + 3) & -4;
 551     saved_ind = ind;
 552     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 553 #ifdef TCC_TARGET_PE
 554     if (v >= 4096) {
 555         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 556         oad(0xb8, v); /* mov stacksize, %eax */
 557         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 558         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 559     } else
 560 #endif
 561     {
 562         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 563         o(0xec81);  /* sub esp, stacksize */
 564         gen_le32(v);
 565 #if FUNC_PROLOG_SIZE == 10
 566         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 567 #endif
 568     }
 569     ind = saved_ind;
 570 }
 571
 572 /* generate a jump to a label */
 573 int gjmp(int t)
 574 {
 575     return psym(0xe9, t);
 576 }
 577
 578 /* generate a jump to a fixed address */
 579 void gjmp_addr(int a)
 580 {
 581     int r;
 582     r = a - ind - 2;
 583     if (r == (char)r) {
 584         g(0xeb);
 585         g(r);
 586     } else {
 587         oad(0xe9, a - ind - 5);
 588     }
 589 }
 590
 591 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 592 int gtst(int inv, int t)
 593 {
 594     int v, *p;
 595
 596     v = vtop->r & VT_VALMASK;
 597     if (v == VT_CMP) {
 598         /* fast case : can jump directly since flags are set */
 599         g(0x0f);
 600         t = psym((vtop->c.i - 16) ^ inv, t);
 601     } else if (v == VT_JMP || v == VT_JMPI) {
 602         /* && or || optimization */
 603         if ((v & 1) == inv) {
 604             /* insert vtop->c jump list in t */
 605             p = &vtop->c.i;
 606             while (*p != 0)
 607                 p = (int *)(cur_text_section->data + *p);
 608             *p = t;
 609             t = vtop->c.i;
 610         } else {
 611             t = gjmp(t);
 612             gsym(vtop->c.i);
 613         }
 614     } else {
 615         if (is_float(vtop->type.t) ||
 616             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 617             vpushi(0);
 618             gen_op(TOK_NE);
 619         }
 620         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 621             /* constant jmp optimization */
 622             if ((vtop->c.i != 0) != inv)
 623                 t = gjmp(t);
 624         } else {
 625             v = gv(RC_INT);
 626             o(0x85);
 627             o(0xc0 + v * 9);
 628             g(0x0f);
 629             t = psym(0x85 ^ inv, t);
 630         }
 631     }
 632     vtop--;
 633     return t;
 634 }
 635
 636 /* generate an integer binary operation */
 637 void gen_opi(int op)
 638 {
 639     int r, fr, opc, c;
 640
 641     switch(op) {
 642     case '+':
 643     case TOK_ADDC1: /* add with carry generation */
 644         opc = 0;
 645     gen_op8:
 646         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 647             /* constant case */
 648             vswap();
 649             r = gv(RC_INT);
 650             vswap();
 651             c = vtop->c.i;
 652             if (c == (char)c) {
 653                 /* XXX: generate inc and dec for smaller code ? */
 654                 o(0x83);
 655                 o(0xc0 | (opc << 3) | r);
 656                 g(c);
 657             } else {
 658                 o(0x81);
 659                 oad(0xc0 | (opc << 3) | r, c);
 660             }
 661         } else {
 662             gv2(RC_INT, RC_INT);
 663             r = vtop[-1].r;
 664             fr = vtop[0].r;
 665             o((opc << 3) | 0x01);
 666             o(0xc0 + r + fr * 8);
 667         }
 668         vtop--;
 669         if (op >= TOK_ULT && op <= TOK_GT) {
 670             vtop->r = VT_CMP;
 671             vtop->c.i = op;
 672         }
 673         break;
 674     case '-':
 675     case TOK_SUBC1: /* sub with carry generation */
 676         opc = 5;
 677         goto gen_op8;
 678     case TOK_ADDC2: /* add with carry use */
 679         opc = 2;
 680         goto gen_op8;
 681     case TOK_SUBC2: /* sub with carry use */
 682         opc = 3;
 683         goto gen_op8;
 684     case '&':
 685         opc = 4;
 686         goto gen_op8;
 687     case '^':
 688         opc = 6;
 689         goto gen_op8;
 690     case '|':
 691         opc = 1;
 692         goto gen_op8;
 693     case '*':
 694         gv2(RC_INT, RC_INT);
 695         r = vtop[-1].r;
 696         fr = vtop[0].r;
 697         vtop--;
 698         o(0xaf0f); /* imul fr, r */
 699         o(0xc0 + fr + r * 8);
 700         break;
 701     case TOK_SHL:
 702         opc = 4;
 703         goto gen_shift;
 704     case TOK_SHR:
 705         opc = 5;
 706         goto gen_shift;
 707     case TOK_SAR:
 708         opc = 7;
 709     gen_shift:
 710         opc = 0xc0 | (opc << 3);
 711         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 712             /* constant case */
 713             vswap();
 714             r = gv(RC_INT);
 715             vswap();
 716             c = vtop->c.i & 0x1f;
 717             o(0xc1); /* shl/shr/sar $xxx, r */
 718             o(opc | r);
 719             g(c);
 720         } else {
 721             /* we generate the shift in ecx */
 722             gv2(RC_INT, RC_ECX);
 723             r = vtop[-1].r;
 724             o(0xd3); /* shl/shr/sar %cl, r */
 725             o(opc | r);
 726         }
 727         vtop--;
 728         break;
 729     case '/':
 730     case TOK_UDIV:
 731     case TOK_PDIV:
 732     case '%':
 733     case TOK_UMOD:
 734     case TOK_UMULL:
 735         /* first operand must be in eax */
 736         /* XXX: need better constraint for second operand */
 737         gv2(RC_EAX, RC_ECX);
 738         r = vtop[-1].r;
 739         fr = vtop[0].r;
 740         vtop--;
 741         save_reg(TREG_EDX);
 742         if (op == TOK_UMULL) {
 743             o(0xf7); /* mul fr */
 744             o(0xe0 + fr);
 745             vtop->r2 = TREG_EDX;
 746             r = TREG_EAX;
 747         } else {
 748             if (op == TOK_UDIV || op == TOK_UMOD) {
 749                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 750                 o(0xf0 + fr);
 751             } else {
 752                 o(0xf799); /* cltd, idiv fr, %eax */
 753                 o(0xf8 + fr);
 754             }
 755             if (op == '%' || op == TOK_UMOD)
 756                 r = TREG_EDX;
 757             else
 758                 r = TREG_EAX;
 759         }
 760         vtop->r = r;
 761         break;
 762     default:
 763         opc = 7;
 764         goto gen_op8;
 765     }
 766 }
 767
 768 /* generate a floating point operation 'v = t1 op t2' instruction. The
 769    two operands are guaranted to have the same floating point type */
 770 /* XXX: need to use ST1 too */
 771 void gen_opf(int op)
 772 {
 773     int a, ft, fc, swapped, r;
 774
 775     /* convert constants to memory references */
 776     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 777         vswap();
 778         gv(RC_FLOAT);
 779         vswap();
 780     }
 781     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 782         gv(RC_FLOAT);
 783
 784     /* must put at least one value in the floating point register */
 785     if ((vtop[-1].r & VT_LVAL) &&
 786         (vtop[0].r & VT_LVAL)) {
 787         vswap();
 788         gv(RC_FLOAT);
 789         vswap();
 790     }
 791     swapped = 0;
 792     /* swap the stack if needed so that t1 is the register and t2 is
 793        the memory reference */
 794     if (vtop[-1].r & VT_LVAL) {
 795         vswap();
 796         swapped = 1;
 797     }
 798     if (op >= TOK_ULT && op <= TOK_GT) {
 799         /* load on stack second operand */
 800         load(TREG_ST0, vtop);
 801         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 802         if (op == TOK_GE || op == TOK_GT)
 803             swapped = !swapped;
 804         else if (op == TOK_EQ || op == TOK_NE)
 805             swapped = 0;
 806         if (swapped)
 807             o(0xc9d9); /* fxch %st(1) */
 808         o(0xe9da); /* fucompp */
 809         o(0xe0df); /* fnstsw %ax */
 810         if (op == TOK_EQ) {
 811             o(0x45e480); /* and $0x45, %ah */
 812             o(0x40fC80); /* cmp $0x40, %ah */
 813         } else if (op == TOK_NE) {
 814             o(0x45e480); /* and $0x45, %ah */
 815             o(0x40f480); /* xor $0x40, %ah */
 816             op = TOK_NE;
 817         } else if (op == TOK_GE || op == TOK_LE) {
 818             o(0x05c4f6); /* test $0x05, %ah */
 819             op = TOK_EQ;
 820         } else {
 821             o(0x45c4f6); /* test $0x45, %ah */
 822             op = TOK_EQ;
 823         }
 824         vtop--;
 825         vtop->r = VT_CMP;
 826         vtop->c.i = op;
 827     } else {
 828         /* no memory reference possible for long double operations */
 829         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 830             load(TREG_ST0, vtop);
 831             swapped = !swapped;
 832         }
 833
 834         switch(op) {
 835         default:
 836         case '+':
 837             a = 0;
 838             break;
 839         case '-':
 840             a = 4;
 841             if (swapped)
 842                 a++;
 843             break;
 844         case '*':
 845             a = 1;
 846             break;
 847         case '/':
 848             a = 6;
 849             if (swapped)
 850                 a++;
 851             break;
 852         }
 853         ft = vtop->type.t;
 854         fc = vtop->c.ul;
 855         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 856             o(0xde); /* fxxxp %st, %st(1) */
 857             o(0xc1 + (a << 3));
 858         } else {
 859             /* if saved lvalue, then we must reload it */
 860             r = vtop->r;
 861             if ((r & VT_VALMASK) == VT_LLOCAL) {
 862                 SValue v1;
 863                 r = get_reg(RC_INT);
 864                 v1.type.t = VT_INT;
 865                 v1.r = VT_LOCAL | VT_LVAL;
 866                 v1.c.ul = fc;
 867                 load(r, &v1);
 868                 fc = 0;
 869             }
 870
 871             if ((ft & VT_BTYPE) == VT_DOUBLE)
 872                 o(0xdc);
 873             else
 874                 o(0xd8);
 875             gen_modrm(a, r, vtop->sym, fc);
 876         }
 877         vtop--;
 878     }
 879 }
 880
 881 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 882    and 'long long' cases. */
 883 void gen_cvt_itof(int t)
 884 {
 885     save_reg(TREG_ST0);
 886     gv(RC_INT);
 887     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 888         /* signed long long to float/double/long double (unsigned case
 889            is handled generically) */
 890         o(0x50 + vtop->r2); /* push r2 */
 891         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 892         o(0x242cdf); /* fildll (%esp) */
 893         o(0x08c483); /* add $8, %esp */
 894     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 895                (VT_INT | VT_UNSIGNED)) {
 896         /* unsigned int to float/double/long double */
 897         o(0x6a); /* push $0 */
 898         g(0x00);
 899         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 900         o(0x242cdf); /* fildll (%esp) */
 901         o(0x08c483); /* add $8, %esp */
 902     } else {
 903         /* int to float/double/long double */
 904         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 905         o(0x2404db); /* fildl (%esp) */
 906         o(0x04c483); /* add $4, %esp */
 907     }
 908     vtop->r = TREG_ST0;
 909 }
 910
 911 /* convert fp to int 't' type */
 912 /* XXX: handle long long case */
 913 void gen_cvt_ftoi(int t)
 914 {
 915     int r, r2, size;
 916     Sym *sym;
 917     CType ushort_type;
 918
 919     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 920
 921     gv(RC_FLOAT);
 922     if (t != VT_INT)
 923         size = 8;
 924     else
 925         size = 4;
 926
 927     o(0x2dd9); /* ldcw xxx */
 928     sym = external_global_sym(TOK___tcc_int_fpu_control,
 929                               &ushort_type, VT_LVAL);
 930     greloc(cur_text_section, sym,
 931            ind, R_386_32);
 932     gen_le32(0);
 933
 934     oad(0xec81, size); /* sub $xxx, %esp */
 935     if (size == 4)
 936         o(0x1cdb); /* fistpl */
 937     else
 938         o(0x3cdf); /* fistpll */
 939     o(0x24);
 940     o(0x2dd9); /* ldcw xxx */
 941     sym = external_global_sym(TOK___tcc_fpu_control,
 942                               &ushort_type, VT_LVAL);
 943     greloc(cur_text_section, sym,
 944            ind, R_386_32);
 945     gen_le32(0);
 946
 947     r = get_reg(RC_INT);
 948     o(0x58 + r); /* pop r */
 949     if (size == 8) {
 950         if (t == VT_LLONG) {
 951             vtop->r = r; /* mark reg as used */
 952             r2 = get_reg(RC_INT);
 953             o(0x58 + r2); /* pop r2 */
 954             vtop->r2 = r2;
 955         } else {
 956             o(0x04c483); /* add $4, %esp */
 957         }
 958     }
 959     vtop->r = r;
 960 }
 961
 962 /* convert from one floating point type to another */
 963 void gen_cvt_ftof(int t)
 964 {
 965     /* all we have to do on i386 is to put the float in a register */
 966     gv(RC_FLOAT);
 967 }
 968
 969 /* computed goto support */
 970 void ggoto(void)
 971 {
 972     gcall_or_jmp(1);
 973     vtop--;
 974 }
 975
 976 /* bound check support functions */
 977 #ifdef CONFIG_TCC_BCHECK
 978
 979 /* generate a bounded pointer addition */
 980 void gen_bounded_ptr_add(void)
 981 {
 982     Sym *sym;
 983
 984     /* prepare fast i386 function call (args in eax and edx) */
 985     gv2(RC_EAX, RC_EDX);
 986     /* save all temporary registers */
 987     vtop -= 2;
 988     save_regs(0);
 989     /* do a fast function call */
 990     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
 991     greloc(cur_text_section, sym,
 992            ind + 1, R_386_PC32);
 993     oad(0xe8, -4);
 994     /* returned pointer is in eax */
 995     vtop++;
 996     vtop->r = TREG_EAX | VT_BOUNDED;
 997     /* address of bounding function call point */
 998     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 999 }
1000
1001 /* patch pointer addition in vtop so that pointer dereferencing is
1002    also tested */
1003 void gen_bounded_ptr_deref(void)
1004 {
1005     int func;
1006     int size, align;
1007     Elf32_Rel *rel;
1008     Sym *sym;
1009
1010     size = 0;
1011     /* XXX: put that code in generic part of tcc */
1012     if (!is_float(vtop->type.t)) {
1013         if (vtop->r & VT_LVAL_BYTE)
1014             size = 1;
1015         else if (vtop->r & VT_LVAL_SHORT)
1016             size = 2;
1017     }
1018     if (!size)
1019         size = type_size(&vtop->type, &align);
1020     switch(size) {
1021     case  1: func = TOK___bound_ptr_indir1; break;
1022     case  2: func = TOK___bound_ptr_indir2; break;
1023     case  4: func = TOK___bound_ptr_indir4; break;
1024     case  8: func = TOK___bound_ptr_indir8; break;
1025     case 12: func = TOK___bound_ptr_indir12; break;
1026     case 16: func = TOK___bound_ptr_indir16; break;
1027     default:
1028         error("unhandled size when derefencing bounded pointer");
1029         func = 0;
1030         break;
1031     }
1032
1033     /* patch relocation */
1034     /* XXX: find a better solution ? */
1035     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1036     sym = external_global_sym(func, &func_old_type, 0);
1037     if (!sym->c)
1038         put_extern_sym(sym, NULL, 0, 0);
1039     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1040 }
1041 #endif
1042
1043 /* end of X86 code generator */
1044 /*************************************************************/
1045