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_le16(int v)
 113 {
 114     g(v);
 115     g(v >> 8);
 116 }
 117
 118 void gen_le32(int c)
 119 {
 120     g(c);
 121     g(c >> 8);
 122     g(c >> 16);
 123     g(c >> 24);
 124 }
 125
 126 /* output a symbol and patch all calls to it */
 127 void gsym_addr(int t, int a)
 128 {
 129     int n, *ptr;
 130     while (t) {
 131         ptr = (int *)(cur_text_section->data + t);
 132         n = *ptr; /* next value */
 133         *ptr = a - t - 4;
 134         t = n;
 135     }
 136 }
 137
 138 void gsym(int t)
 139 {
 140     gsym_addr(t, ind);
 141 }
 142
 143 /* psym is used to put an instruction with a data field which is a
 144    reference to a symbol. It is in fact the same as oad ! */
 145 #define psym oad
 146
 147 /* instruction + 4 bytes data. Return the address of the data */
 148 static int oad(int c, int s)
 149 {
 150     int ind1;
 151
 152     o(c);
 153     ind1 = ind + 4;
 154     if (ind1 > cur_text_section->data_allocated)
 155         section_realloc(cur_text_section, ind1);
 156     *(int *)(cur_text_section->data + ind) = s;
 157     s = ind;
 158     ind = ind1;
 159     return s;
 160 }
 161
 162 /* output constant with relocation if 'r & VT_SYM' is true */
 163 static void gen_addr32(int r, Sym *sym, int c)
 164 {
 165     if (r & VT_SYM)
 166         greloc(cur_text_section, sym, ind, R_386_32);
 167     gen_le32(c);
 168 }
 169
 170 static void gen_addrpc32(int r, Sym *sym, int c)
 171 {
 172     if (r & VT_SYM)
 173         greloc(cur_text_section, sym, ind, R_386_PC32);
 174     gen_le32(c - 4);
 175 }
 176
 177 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 178    opcode bits */
 179 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 180 {
 181     op_reg = op_reg << 3;
 182     if ((r & VT_VALMASK) == VT_CONST) {
 183         /* constant memory reference */
 184         o(0x05 | op_reg);
 185         gen_addr32(r, sym, c);
 186     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 187         /* currently, we use only ebp as base */
 188         if (c == (char)c) {
 189             /* short reference */
 190             o(0x45 | op_reg);
 191             g(c);
 192         } else {
 193             oad(0x85 | op_reg, c);
 194         }
 195     } else {
 196         g(0x00 | op_reg | (r & VT_VALMASK));
 197     }
 198 }
 199
 200 /* load 'r' from value 'sv' */
 201 void load(int r, SValue *sv)
 202 {
 203     int v, t, ft, fc, fr;
 204     SValue v1;
 205
 206 #ifdef TCC_TARGET_PE
 207     if (pe_dllimport(r, sv, load))
 208         return;
 209 #endif
 210     fr = sv->r;
 211     ft = sv->type.t;
 212     fc = sv->c.ul;
 213
 214     v = fr & VT_VALMASK;
 215     if (fr & VT_LVAL) {
 216         if (v == VT_LLOCAL) {
 217             v1.type.t = VT_INT;
 218             v1.r = VT_LOCAL | VT_LVAL;
 219             v1.c.ul = fc;
 220             load(r, &v1);
 221             fr = r;
 222         }
 223         if ((ft & VT_BTYPE) == VT_FLOAT) {
 224             o(0xd9); /* flds */
 225             r = 0;
 226         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 227             o(0xdd); /* fldl */
 228             r = 0;
 229         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 230             o(0xdb); /* fldt */
 231             r = 5;
 232         } else if ((ft & VT_TYPE) == VT_BYTE) {
 233             o(0xbe0f);   /* movsbl */
 234         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 235             o(0xb60f);   /* movzbl */
 236         } else if ((ft & VT_TYPE) == VT_SHORT) {
 237             o(0xbf0f);   /* movswl */
 238         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 239             o(0xb70f);   /* movzwl */
 240         } else {
 241             o(0x8b);     /* movl */
 242         }
 243         gen_modrm(r, fr, sv->sym, fc);
 244     } else {
 245         if (v == VT_CONST) {
 246             o(0xb8 + r); /* mov $xx, r */
 247             gen_addr32(fr, sv->sym, fc);
 248         } else if (v == VT_LOCAL) {
 249             o(0x8d); /* lea xxx(%ebp), r */
 250             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 251         } else if (v == VT_CMP) {
 252             oad(0xb8 + r, 0); /* mov $0, r */
 253             o(0x0f); /* setxx %br */
 254             o(fc);
 255             o(0xc0 + r);
 256         } else if (v == VT_JMP || v == VT_JMPI) {
 257             t = v & 1;
 258             oad(0xb8 + r, t); /* mov $1, r */
 259             o(0x05eb); /* jmp after */
 260             gsym(fc);
 261             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 262         } else if (v != r) {
 263             o(0x89);
 264             o(0xc0 + r + v * 8); /* mov v, r */
 265         }
 266     }
 267 }
 268
 269 /* store register 'r' in lvalue 'v' */
 270 void store(int r, SValue *v)
 271 {
 272     int fr, bt, ft, fc;
 273
 274 #ifdef TCC_TARGET_PE
 275     if (pe_dllimport(r, v, store))
 276         return;
 277 #endif
 278     ft = v->type.t;
 279     fc = v->c.ul;
 280     fr = v->r & VT_VALMASK;
 281     bt = ft & VT_BTYPE;
 282     /* XXX: incorrect if float reg to reg */
 283     if (bt == VT_FLOAT) {
 284         o(0xd9); /* fsts */
 285         r = 2;
 286     } else if (bt == VT_DOUBLE) {
 287         o(0xdd); /* fstpl */
 288         r = 2;
 289     } else if (bt == VT_LDOUBLE) {
 290         o(0xc0d9); /* fld %st(0) */
 291         o(0xdb); /* fstpt */
 292         r = 7;
 293     } else {
 294         if (bt == VT_SHORT)
 295             o(0x66);
 296         if (bt == VT_BYTE || bt == VT_BOOL)
 297             o(0x88);
 298         else
 299             o(0x89);
 300     }
 301     if (fr == VT_CONST ||
 302         fr == VT_LOCAL ||
 303         (v->r & VT_LVAL)) {
 304         gen_modrm(r, v->r, v->sym, fc);
 305     } else if (fr != r) {
 306         o(0xc0 + fr + r * 8); /* mov r, fr */
 307     }
 308 }
 309
 310 static void gadd_sp(int val)
 311 {
 312     if (val == (char)val) {
 313         o(0xc483);
 314         g(val);
 315     } else {
 316         oad(0xc481, val); /* add $xxx, %esp */
 317     }
 318 }
 319
 320 /* 'is_jmp' is '1' if it is a jump */
 321 static void gcall_or_jmp(int is_jmp)
 322 {
 323     int r;
 324     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 325         /* constant case */
 326         if (vtop->r & VT_SYM) {
 327             /* relocation case */
 328             greloc(cur_text_section, vtop->sym,
 329                    ind + 1, R_386_PC32);
 330         } else {
 331             /* put an empty PC32 relocation */
 332             put_elf_reloc(symtab_section, cur_text_section,
 333                           ind + 1, R_386_PC32, 0);
 334         }
 335         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 336     } else {
 337         /* otherwise, indirect call */
 338         r = gv(RC_INT);
 339         o(0xff); /* call/jmp *r */
 340         o(0xd0 + r + (is_jmp << 4));
 341     }
 342 }
 343
 344 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 345 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 346
 347 /* Generate function call. The function address is pushed first, then
 348    all the parameters in call order. This functions pops all the
 349    parameters and the function address. */
 350 void gfunc_call(int nb_args)
 351 {
 352     int size, align, r, args_size, i, func_call;
 353     Sym *func_sym;
 354
 355     args_size = 0;
 356     for(i = 0;i < nb_args; i++) {
 357         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 358             size = type_size(&vtop->type, &align);
 359             /* align to stack align size */
 360             size = (size + 3) & ~3;
 361             /* allocate the necessary size on stack */
 362             oad(0xec81, size); /* sub $xxx, %esp */
 363             /* generate structure store */
 364             r = get_reg(RC_INT);
 365             o(0x89); /* mov %esp, r */
 366             o(0xe0 + r);
 367             vset(&vtop->type, r | VT_LVAL, 0);
 368             vswap();
 369             vstore();
 370             args_size += size;
 371         } else if (is_float(vtop->type.t)) {
 372             gv(RC_FLOAT); /* only one float register */
 373             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 374                 size = 4;
 375             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 376                 size = 8;
 377             else
 378                 size = 12;
 379             oad(0xec81, size); /* sub $xxx, %esp */
 380             if (size == 12)
 381                 o(0x7cdb);
 382             else
 383                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 384             g(0x24);
 385             g(0x00);
 386             args_size += size;
 387         } else {
 388             /* simple type (currently always same size) */
 389             /* XXX: implicit cast ? */
 390             r = gv(RC_INT);
 391             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 392                 size = 8;
 393                 o(0x50 + vtop->r2); /* push r */
 394             } else {
 395                 size = 4;
 396             }
 397             o(0x50 + r); /* push r */
 398             args_size += size;
 399         }
 400         vtop--;
 401     }
 402     save_regs(0); /* save used temporary registers */
 403     func_sym = vtop->type.ref;
 404     func_call = FUNC_CALL(func_sym->r);
 405     /* fast call case */
 406     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 407         func_call == FUNC_FASTCALLW) {
 408         int fastcall_nb_regs;
 409         uint8_t *fastcall_regs_ptr;
 410         if (func_call == FUNC_FASTCALLW) {
 411             fastcall_regs_ptr = fastcallw_regs;
 412             fastcall_nb_regs = 2;
 413         } else {
 414             fastcall_regs_ptr = fastcall_regs;
 415             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 416         }
 417         for(i = 0;i < fastcall_nb_regs; i++) {
 418             if (args_size <= 0)
 419                 break;
 420             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 421             /* XXX: incorrect for struct/floats */
 422             args_size -= 4;
 423         }
 424     }
 425     gcall_or_jmp(0);
 426
 427 #ifdef TCC_TARGET_PE
 428     if ((func_sym->type.t & VT_BTYPE) == VT_STRUCT)
 429         args_size -= 4;
 430 #endif
 431     if (args_size && func_call != FUNC_STDCALL)
 432         gadd_sp(args_size);
 433     vtop--;
 434 }
 435
 436 #ifdef TCC_TARGET_PE
 437 #define FUNC_PROLOG_SIZE 10
 438 #else
 439 #define FUNC_PROLOG_SIZE 9
 440 #endif
 441
 442 /* generate function prolog of type 't' */
 443 void gfunc_prolog(CType *func_type)
 444 {
 445     int addr, align, size, func_call, fastcall_nb_regs;
 446     int param_index, param_addr;
 447     uint8_t *fastcall_regs_ptr;
 448     Sym *sym;
 449     CType *type;
 450
 451     sym = func_type->ref;
 452     func_call = FUNC_CALL(sym->r);
 453     addr = 8;
 454     loc = 0;
 455     func_vc = 0;
 456
 457     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 458         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 459         fastcall_regs_ptr = fastcall_regs;
 460     } else if (func_call == FUNC_FASTCALLW) {
 461         fastcall_nb_regs = 2;
 462         fastcall_regs_ptr = fastcallw_regs;
 463     } else {
 464         fastcall_nb_regs = 0;
 465         fastcall_regs_ptr = NULL;
 466     }
 467     param_index = 0;
 468
 469     ind += FUNC_PROLOG_SIZE;
 470     func_sub_sp_offset = ind;
 471     /* if the function returns a structure, then add an
 472        implicit pointer parameter */
 473     func_vt = sym->type;
 474     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 475         /* XXX: fastcall case ? */
 476         func_vc = addr;
 477         addr += 4;
 478         param_index++;
 479     }
 480     /* define parameters */
 481     while ((sym = sym->next) != NULL) {
 482         type = &sym->type;
 483         size = type_size(type, &align);
 484         size = (size + 3) & ~3;
 485 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 486         /* structs are passed as pointer */
 487         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 488             size = 4;
 489         }
 490 #endif
 491         if (param_index < fastcall_nb_regs) {
 492             /* save FASTCALL register */
 493             loc -= 4;
 494             o(0x89);     /* movl */
 495             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 496             param_addr = loc;
 497         } else {
 498             param_addr = addr;
 499             addr += size;
 500         }
 501         sym_push(sym->v & ~SYM_FIELD, type,
 502                  VT_LOCAL | lvalue_type(type->t), param_addr);
 503         param_index++;
 504     }
 505     func_ret_sub = 0;
 506     /* pascal type call ? */
 507     if (func_call == FUNC_STDCALL)
 508         func_ret_sub = addr - 8;
 509 #ifdef TCC_TARGET_PE
 510     else if (func_vc)
 511         func_ret_sub = 4;
 512 #endif
 513
 514     /* leave some room for bound checking code */
 515     if (tcc_state->do_bounds_check) {
 516         oad(0xb8, 0); /* lbound section pointer */
 517         oad(0xb8, 0); /* call to function */
 518         func_bound_offset = lbounds_section->data_offset;
 519     }
 520 }
 521
 522 /* generate function epilog */
 523 void gfunc_epilog(void)
 524 {
 525     int v, saved_ind;
 526
 527 #ifdef CONFIG_TCC_BCHECK
 528     if (tcc_state->do_bounds_check
 529      && func_bound_offset != lbounds_section->data_offset) {
 530         int saved_ind;
 531         int *bounds_ptr;
 532         Sym *sym, *sym_data;
 533         /* add end of table info */
 534         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 535         *bounds_ptr = 0;
 536         /* generate bound local allocation */
 537         saved_ind = ind;
 538         ind = func_sub_sp_offset;
 539         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 540                                func_bound_offset, lbounds_section->data_offset);
 541         greloc(cur_text_section, sym_data,
 542                ind + 1, R_386_32);
 543         oad(0xb8, 0); /* mov %eax, xxx */
 544         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 545         greloc(cur_text_section, sym,
 546                ind + 1, R_386_PC32);
 547         oad(0xe8, -4);
 548         ind = saved_ind;
 549         /* generate bound check local freeing */
 550         o(0x5250); /* save returned value, if any */
 551         greloc(cur_text_section, sym_data,
 552                ind + 1, R_386_32);
 553         oad(0xb8, 0); /* mov %eax, xxx */
 554         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 555         greloc(cur_text_section, sym,
 556                ind + 1, R_386_PC32);
 557         oad(0xe8, -4);
 558         o(0x585a); /* restore returned value, if any */
 559     }
 560 #endif
 561     o(0xc9); /* leave */
 562     if (func_ret_sub == 0) {
 563         o(0xc3); /* ret */
 564     } else {
 565         o(0xc2); /* ret n */
 566         g(func_ret_sub);
 567         g(func_ret_sub >> 8);
 568     }
 569     /* align local size to word & save local variables */
 570
 571     v = (-loc + 3) & -4;
 572     saved_ind = ind;
 573     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 574 #ifdef TCC_TARGET_PE
 575     if (v >= 4096) {
 576         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 577         oad(0xb8, v); /* mov stacksize, %eax */
 578         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 579         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 580     } else
 581 #endif
 582     {
 583         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 584         o(0xec81);  /* sub esp, stacksize */
 585         gen_le32(v);
 586 #if FUNC_PROLOG_SIZE == 10
 587         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 588 #endif
 589     }
 590     ind = saved_ind;
 591 }
 592
 593 /* generate a jump to a label */
 594 int gjmp(int t)
 595 {
 596     return psym(0xe9, t);
 597 }
 598
 599 /* generate a jump to a fixed address */
 600 void gjmp_addr(int a)
 601 {
 602     int r;
 603     r = a - ind - 2;
 604     if (r == (char)r) {
 605         g(0xeb);
 606         g(r);
 607     } else {
 608         oad(0xe9, a - ind - 5);
 609     }
 610 }
 611
 612 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 613 int gtst(int inv, int t)
 614 {
 615     int v, *p;
 616
 617     v = vtop->r & VT_VALMASK;
 618     if (v == VT_CMP) {
 619         /* fast case : can jump directly since flags are set */
 620         g(0x0f);
 621         t = psym((vtop->c.i - 16) ^ inv, t);
 622     } else if (v == VT_JMP || v == VT_JMPI) {
 623         /* && or || optimization */
 624         if ((v & 1) == inv) {
 625             /* insert vtop->c jump list in t */
 626             p = &vtop->c.i;
 627             while (*p != 0)
 628                 p = (int *)(cur_text_section->data + *p);
 629             *p = t;
 630             t = vtop->c.i;
 631         } else {
 632             t = gjmp(t);
 633             gsym(vtop->c.i);
 634         }
 635     } else {
 636         if (is_float(vtop->type.t) ||
 637             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 638             vpushi(0);
 639             gen_op(TOK_NE);
 640         }
 641         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 642             /* constant jmp optimization */
 643             if ((vtop->c.i != 0) != inv)
 644                 t = gjmp(t);
 645         } else {
 646             v = gv(RC_INT);
 647             o(0x85);
 648             o(0xc0 + v * 9);
 649             g(0x0f);
 650             t = psym(0x85 ^ inv, t);
 651         }
 652     }
 653     vtop--;
 654     return t;
 655 }
 656
 657 /* generate an integer binary operation */
 658 void gen_opi(int op)
 659 {
 660     int r, fr, opc, c;
 661
 662     switch(op) {
 663     case '+':
 664     case TOK_ADDC1: /* add with carry generation */
 665         opc = 0;
 666     gen_op8:
 667         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 668             /* constant case */
 669             vswap();
 670             r = gv(RC_INT);
 671             vswap();
 672             c = vtop->c.i;
 673             if (c == (char)c) {
 674                 /* XXX: generate inc and dec for smaller code ? */
 675                 o(0x83);
 676                 o(0xc0 | (opc << 3) | r);
 677                 g(c);
 678             } else {
 679                 o(0x81);
 680                 oad(0xc0 | (opc << 3) | r, c);
 681             }
 682         } else {
 683             gv2(RC_INT, RC_INT);
 684             r = vtop[-1].r;
 685             fr = vtop[0].r;
 686             o((opc << 3) | 0x01);
 687             o(0xc0 + r + fr * 8);
 688         }
 689         vtop--;
 690         if (op >= TOK_ULT && op <= TOK_GT) {
 691             vtop->r = VT_CMP;
 692             vtop->c.i = op;
 693         }
 694         break;
 695     case '-':
 696     case TOK_SUBC1: /* sub with carry generation */
 697         opc = 5;
 698         goto gen_op8;
 699     case TOK_ADDC2: /* add with carry use */
 700         opc = 2;
 701         goto gen_op8;
 702     case TOK_SUBC2: /* sub with carry use */
 703         opc = 3;
 704         goto gen_op8;
 705     case '&':
 706         opc = 4;
 707         goto gen_op8;
 708     case '^':
 709         opc = 6;
 710         goto gen_op8;
 711     case '|':
 712         opc = 1;
 713         goto gen_op8;
 714     case '*':
 715         gv2(RC_INT, RC_INT);
 716         r = vtop[-1].r;
 717         fr = vtop[0].r;
 718         vtop--;
 719         o(0xaf0f); /* imul fr, r */
 720         o(0xc0 + fr + r * 8);
 721         break;
 722     case TOK_SHL:
 723         opc = 4;
 724         goto gen_shift;
 725     case TOK_SHR:
 726         opc = 5;
 727         goto gen_shift;
 728     case TOK_SAR:
 729         opc = 7;
 730     gen_shift:
 731         opc = 0xc0 | (opc << 3);
 732         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 733             /* constant case */
 734             vswap();
 735             r = gv(RC_INT);
 736             vswap();
 737             c = vtop->c.i & 0x1f;
 738             o(0xc1); /* shl/shr/sar $xxx, r */
 739             o(opc | r);
 740             g(c);
 741         } else {
 742             /* we generate the shift in ecx */
 743             gv2(RC_INT, RC_ECX);
 744             r = vtop[-1].r;
 745             o(0xd3); /* shl/shr/sar %cl, r */
 746             o(opc | r);
 747         }
 748         vtop--;
 749         break;
 750     case '/':
 751     case TOK_UDIV:
 752     case TOK_PDIV:
 753     case '%':
 754     case TOK_UMOD:
 755     case TOK_UMULL:
 756         /* first operand must be in eax */
 757         /* XXX: need better constraint for second operand */
 758         gv2(RC_EAX, RC_ECX);
 759         r = vtop[-1].r;
 760         fr = vtop[0].r;
 761         vtop--;
 762         save_reg(TREG_EDX);
 763         if (op == TOK_UMULL) {
 764             o(0xf7); /* mul fr */
 765             o(0xe0 + fr);
 766             vtop->r2 = TREG_EDX;
 767             r = TREG_EAX;
 768         } else {
 769             if (op == TOK_UDIV || op == TOK_UMOD) {
 770                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 771                 o(0xf0 + fr);
 772             } else {
 773                 o(0xf799); /* cltd, idiv fr, %eax */
 774                 o(0xf8 + fr);
 775             }
 776             if (op == '%' || op == TOK_UMOD)
 777                 r = TREG_EDX;
 778             else
 779                 r = TREG_EAX;
 780         }
 781         vtop->r = r;
 782         break;
 783     default:
 784         opc = 7;
 785         goto gen_op8;
 786     }
 787 }
 788
 789 /* generate a floating point operation 'v = t1 op t2' instruction. The
 790    two operands are guaranted to have the same floating point type */
 791 /* XXX: need to use ST1 too */
 792 void gen_opf(int op)
 793 {
 794     int a, ft, fc, swapped, r;
 795
 796     /* convert constants to memory references */
 797     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 798         vswap();
 799         gv(RC_FLOAT);
 800         vswap();
 801     }
 802     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 803         gv(RC_FLOAT);
 804
 805     /* must put at least one value in the floating point register */
 806     if ((vtop[-1].r & VT_LVAL) &&
 807         (vtop[0].r & VT_LVAL)) {
 808         vswap();
 809         gv(RC_FLOAT);
 810         vswap();
 811     }
 812     swapped = 0;
 813     /* swap the stack if needed so that t1 is the register and t2 is
 814        the memory reference */
 815     if (vtop[-1].r & VT_LVAL) {
 816         vswap();
 817         swapped = 1;
 818     }
 819     if (op >= TOK_ULT && op <= TOK_GT) {
 820         /* load on stack second operand */
 821         load(TREG_ST0, vtop);
 822         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 823         if (op == TOK_GE || op == TOK_GT)
 824             swapped = !swapped;
 825         else if (op == TOK_EQ || op == TOK_NE)
 826             swapped = 0;
 827         if (swapped)
 828             o(0xc9d9); /* fxch %st(1) */
 829         o(0xe9da); /* fucompp */
 830         o(0xe0df); /* fnstsw %ax */
 831         if (op == TOK_EQ) {
 832             o(0x45e480); /* and $0x45, %ah */
 833             o(0x40fC80); /* cmp $0x40, %ah */
 834         } else if (op == TOK_NE) {
 835             o(0x45e480); /* and $0x45, %ah */
 836             o(0x40f480); /* xor $0x40, %ah */
 837             op = TOK_NE;
 838         } else if (op == TOK_GE || op == TOK_LE) {
 839             o(0x05c4f6); /* test $0x05, %ah */
 840             op = TOK_EQ;
 841         } else {
 842             o(0x45c4f6); /* test $0x45, %ah */
 843             op = TOK_EQ;
 844         }
 845         vtop--;
 846         vtop->r = VT_CMP;
 847         vtop->c.i = op;
 848     } else {
 849         /* no memory reference possible for long double operations */
 850         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 851             load(TREG_ST0, vtop);
 852             swapped = !swapped;
 853         }
 854
 855         switch(op) {
 856         default:
 857         case '+':
 858             a = 0;
 859             break;
 860         case '-':
 861             a = 4;
 862             if (swapped)
 863                 a++;
 864             break;
 865         case '*':
 866             a = 1;
 867             break;
 868         case '/':
 869             a = 6;
 870             if (swapped)
 871                 a++;
 872             break;
 873         }
 874         ft = vtop->type.t;
 875         fc = vtop->c.ul;
 876         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 877             o(0xde); /* fxxxp %st, %st(1) */
 878             o(0xc1 + (a << 3));
 879         } else {
 880             /* if saved lvalue, then we must reload it */
 881             r = vtop->r;
 882             if ((r & VT_VALMASK) == VT_LLOCAL) {
 883                 SValue v1;
 884                 r = get_reg(RC_INT);
 885                 v1.type.t = VT_INT;
 886                 v1.r = VT_LOCAL | VT_LVAL;
 887                 v1.c.ul = fc;
 888                 load(r, &v1);
 889                 fc = 0;
 890             }
 891
 892             if ((ft & VT_BTYPE) == VT_DOUBLE)
 893                 o(0xdc);
 894             else
 895                 o(0xd8);
 896             gen_modrm(a, r, vtop->sym, fc);
 897         }
 898         vtop--;
 899     }
 900 }
 901
 902 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 903    and 'long long' cases. */
 904 void gen_cvt_itof(int t)
 905 {
 906     save_reg(TREG_ST0);
 907     gv(RC_INT);
 908     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 909         /* signed long long to float/double/long double (unsigned case
 910            is handled generically) */
 911         o(0x50 + vtop->r2); /* push r2 */
 912         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 913         o(0x242cdf); /* fildll (%esp) */
 914         o(0x08c483); /* add $8, %esp */
 915     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 916                (VT_INT | VT_UNSIGNED)) {
 917         /* unsigned int to float/double/long double */
 918         o(0x6a); /* push $0 */
 919         g(0x00);
 920         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 921         o(0x242cdf); /* fildll (%esp) */
 922         o(0x08c483); /* add $8, %esp */
 923     } else {
 924         /* int to float/double/long double */
 925         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 926         o(0x2404db); /* fildl (%esp) */
 927         o(0x04c483); /* add $4, %esp */
 928     }
 929     vtop->r = TREG_ST0;
 930 }
 931
 932 /* convert fp to int 't' type */
 933 /* XXX: handle long long case */
 934 void gen_cvt_ftoi(int t)
 935 {
 936     int r, r2, size;
 937     Sym *sym;
 938     CType ushort_type;
 939
 940     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 941
 942     gv(RC_FLOAT);
 943     if (t != VT_INT)
 944         size = 8;
 945     else
 946         size = 4;
 947
 948     o(0x2dd9); /* ldcw xxx */
 949     sym = external_global_sym(TOK___tcc_int_fpu_control,
 950                               &ushort_type, VT_LVAL);
 951     greloc(cur_text_section, sym,
 952            ind, R_386_32);
 953     gen_le32(0);
 954
 955     oad(0xec81, size); /* sub $xxx, %esp */
 956     if (size == 4)
 957         o(0x1cdb); /* fistpl */
 958     else
 959         o(0x3cdf); /* fistpll */
 960     o(0x24);
 961     o(0x2dd9); /* ldcw xxx */
 962     sym = external_global_sym(TOK___tcc_fpu_control,
 963                               &ushort_type, VT_LVAL);
 964     greloc(cur_text_section, sym,
 965            ind, R_386_32);
 966     gen_le32(0);
 967
 968     r = get_reg(RC_INT);
 969     o(0x58 + r); /* pop r */
 970     if (size == 8) {
 971         if (t == VT_LLONG) {
 972             vtop->r = r; /* mark reg as used */
 973             r2 = get_reg(RC_INT);
 974             o(0x58 + r2); /* pop r2 */
 975             vtop->r2 = r2;
 976         } else {
 977             o(0x04c483); /* add $4, %esp */
 978         }
 979     }
 980     vtop->r = r;
 981 }
 982
 983 /* convert from one floating point type to another */
 984 void gen_cvt_ftof(int t)
 985 {
 986     /* all we have to do on i386 is to put the float in a register */
 987     gv(RC_FLOAT);
 988 }
 989
 990 /* computed goto support */
 991 void ggoto(void)
 992 {
 993     gcall_or_jmp(1);
 994     vtop--;
 995 }
 996
 997 /* bound check support functions */
 998 #ifdef CONFIG_TCC_BCHECK
 999
1000 /* generate a bounded pointer addition */
1001 void gen_bounded_ptr_add(void)
1002 {
1003     Sym *sym;
1004
1005     /* prepare fast i386 function call (args in eax and edx) */
1006     gv2(RC_EAX, RC_EDX);
1007     /* save all temporary registers */
1008     vtop -= 2;
1009     save_regs(0);
1010     /* do a fast function call */
1011     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
1012     greloc(cur_text_section, sym,
1013            ind + 1, R_386_PC32);
1014     oad(0xe8, -4);
1015     /* returned pointer is in eax */
1016     vtop++;
1017     vtop->r = TREG_EAX | VT_BOUNDED;
1018     /* address of bounding function call point */
1019     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1020 }
1021
1022 /* patch pointer addition in vtop so that pointer dereferencing is
1023    also tested */
1024 void gen_bounded_ptr_deref(void)
1025 {
1026     int func;
1027     int size, align;
1028     Elf32_Rel *rel;
1029     Sym *sym;
1030
1031     size = 0;
1032     /* XXX: put that code in generic part of tcc */
1033     if (!is_float(vtop->type.t)) {
1034         if (vtop->r & VT_LVAL_BYTE)
1035             size = 1;
1036         else if (vtop->r & VT_LVAL_SHORT)
1037             size = 2;
1038     }
1039     if (!size)
1040         size = type_size(&vtop->type, &align);
1041     switch(size) {
1042     case  1: func = TOK___bound_ptr_indir1; break;
1043     case  2: func = TOK___bound_ptr_indir2; break;
1044     case  4: func = TOK___bound_ptr_indir4; break;
1045     case  8: func = TOK___bound_ptr_indir8; break;
1046     case 12: func = TOK___bound_ptr_indir12; break;
1047     case 16: func = TOK___bound_ptr_indir16; break;
1048     default:
1049         error("unhandled size when derefencing bounded pointer");
1050         func = 0;
1051         break;
1052     }
1053
1054     /* patch relocation */
1055     /* XXX: find a better solution ? */
1056     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1057     sym = external_global_sym(func, &func_old_type, 0);
1058     if (!sym->c)
1059         put_extern_sym(sym, NULL, 0, 0);
1060     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1061 }
1062 #endif
1063
1064 /* end of X86 code generator */
1065 /*************************************************************/
1066