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 #ifdef TCC_TARGET_PE
 201 static void mk_pointer(CType *type);
 202 static void indir(void);
 203
 204 int handle_dllimport(int r, SValue *sv, void (*fn)(int r, SValue *sv))
 205 {
 206     if ((sv->r & (VT_VALMASK|VT_SYM|VT_CONST)) != (VT_SYM|VT_CONST))
 207         return 0;
 208     if (0 == (sv->sym->type.t & VT_IMPORT))
 209         return 0;
 210
 211     printf("import %d %04x %s\n", r, ind, get_tok_str(sv->sym->v, NULL));
 212
 213     sv->sym->type.t &= ~VT_IMPORT;
 214     ++vtop;
 215
 216     *vtop = *sv;
 217     mk_pointer(&vtop->type);
 218     indir();
 219     fn(r, vtop);
 220
 221     --vtop;
 222     sv->sym->type.t |= VT_IMPORT;
 223     return 1;
 224 }
 225 #endif
 226
 227 /* load 'r' from value 'sv' */
 228 void load(int r, SValue *sv)
 229 {
 230     int v, t, ft, fc, fr;
 231     SValue v1;
 232
 233 #ifdef TCC_TARGET_PE
 234     if (handle_dllimport(r, sv, load))
 235         return;
 236 #endif
 237     fr = sv->r;
 238     ft = sv->type.t;
 239     fc = sv->c.ul;
 240
 241     v = fr & VT_VALMASK;
 242     if (fr & VT_LVAL) {
 243         if (v == VT_LLOCAL) {
 244             v1.type.t = VT_INT;
 245             v1.r = VT_LOCAL | VT_LVAL;
 246             v1.c.ul = fc;
 247             load(r, &v1);
 248             fr = r;
 249         }
 250         if ((ft & VT_BTYPE) == VT_FLOAT) {
 251             o(0xd9); /* flds */
 252             r = 0;
 253         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 254             o(0xdd); /* fldl */
 255             r = 0;
 256         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 257             o(0xdb); /* fldt */
 258             r = 5;
 259         } else if ((ft & VT_TYPE) == VT_BYTE) {
 260             o(0xbe0f);   /* movsbl */
 261         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 262             o(0xb60f);   /* movzbl */
 263         } else if ((ft & VT_TYPE) == VT_SHORT) {
 264             o(0xbf0f);   /* movswl */
 265         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 266             o(0xb70f);   /* movzwl */
 267         } else {
 268             o(0x8b);     /* movl */
 269         }
 270         gen_modrm(r, fr, sv->sym, fc);
 271     } else {
 272         if (v == VT_CONST) {
 273             o(0xb8 + r); /* mov $xx, r */
 274             gen_addr32(fr, sv->sym, fc);
 275         } else if (v == VT_LOCAL) {
 276             o(0x8d); /* lea xxx(%ebp), r */
 277             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 278         } else if (v == VT_CMP) {
 279             oad(0xb8 + r, 0); /* mov $0, r */
 280             o(0x0f); /* setxx %br */
 281             o(fc);
 282             o(0xc0 + r);
 283         } else if (v == VT_JMP || v == VT_JMPI) {
 284             t = v & 1;
 285             oad(0xb8 + r, t); /* mov $1, r */
 286             o(0x05eb); /* jmp after */
 287             gsym(fc);
 288             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 289         } else if (v != r) {
 290             o(0x89);
 291             o(0xc0 + r + v * 8); /* mov v, r */
 292         }
 293     }
 294 }
 295
 296 /* store register 'r' in lvalue 'v' */
 297 void store(int r, SValue *v)
 298 {
 299     int fr, bt, ft, fc;
 300
 301 #ifdef TCC_TARGET_PE
 302     if (handle_dllimport(r, v, store))
 303         return;
 304 #endif
 305     ft = v->type.t;
 306     fc = v->c.ul;
 307     fr = v->r & VT_VALMASK;
 308     bt = ft & VT_BTYPE;
 309     /* XXX: incorrect if float reg to reg */
 310     if (bt == VT_FLOAT) {
 311         o(0xd9); /* fsts */
 312         r = 2;
 313     } else if (bt == VT_DOUBLE) {
 314         o(0xdd); /* fstpl */
 315         r = 2;
 316     } else if (bt == VT_LDOUBLE) {
 317         o(0xc0d9); /* fld %st(0) */
 318         o(0xdb); /* fstpt */
 319         r = 7;
 320     } else {
 321         if (bt == VT_SHORT)
 322             o(0x66);
 323         if (bt == VT_BYTE || bt == VT_BOOL)
 324             o(0x88);
 325         else
 326             o(0x89);
 327     }
 328     if (fr == VT_CONST ||
 329         fr == VT_LOCAL ||
 330         (v->r & VT_LVAL)) {
 331         gen_modrm(r, v->r, v->sym, fc);
 332     } else if (fr != r) {
 333         o(0xc0 + fr + r * 8); /* mov r, fr */
 334     }
 335 }
 336
 337 static void gadd_sp(int val)
 338 {
 339     if (val == (char)val) {
 340         o(0xc483);
 341         g(val);
 342     } else {
 343         oad(0xc481, val); /* add $xxx, %esp */
 344     }
 345 }
 346
 347 /* 'is_jmp' is '1' if it is a jump */
 348 static void gcall_or_jmp(int is_jmp)
 349 {
 350     int r;
 351     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 352         /* constant case */
 353         if (vtop->r & VT_SYM) {
 354             /* relocation case */
 355             greloc(cur_text_section, vtop->sym,
 356                    ind + 1, R_386_PC32);
 357         } else {
 358             /* put an empty PC32 relocation */
 359             put_elf_reloc(symtab_section, cur_text_section,
 360                           ind + 1, R_386_PC32, 0);
 361         }
 362         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 363     } else {
 364         /* otherwise, indirect call */
 365         r = gv(RC_INT);
 366         o(0xff); /* call/jmp *r */
 367         o(0xd0 + r + (is_jmp << 4));
 368     }
 369 }
 370
 371 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 372 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 373
 374 /* Generate function call. The function address is pushed first, then
 375    all the parameters in call order. This functions pops all the
 376    parameters and the function address. */
 377 void gfunc_call(int nb_args)
 378 {
 379     int size, align, r, args_size, i, func_call;
 380     Sym *func_sym;
 381
 382     args_size = 0;
 383     for(i = 0;i < nb_args; i++) {
 384         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 385             size = type_size(&vtop->type, &align);
 386             /* align to stack align size */
 387             size = (size + 3) & ~3;
 388             /* allocate the necessary size on stack */
 389             oad(0xec81, size); /* sub $xxx, %esp */
 390             /* generate structure store */
 391             r = get_reg(RC_INT);
 392             o(0x89); /* mov %esp, r */
 393             o(0xe0 + r);
 394             vset(&vtop->type, r | VT_LVAL, 0);
 395             vswap();
 396             vstore();
 397             args_size += size;
 398         } else if (is_float(vtop->type.t)) {
 399             gv(RC_FLOAT); /* only one float register */
 400             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 401                 size = 4;
 402             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 403                 size = 8;
 404             else
 405                 size = 12;
 406             oad(0xec81, size); /* sub $xxx, %esp */
 407             if (size == 12)
 408                 o(0x7cdb);
 409             else
 410                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 411             g(0x24);
 412             g(0x00);
 413             args_size += size;
 414         } else {
 415             /* simple type (currently always same size) */
 416             /* XXX: implicit cast ? */
 417             r = gv(RC_INT);
 418             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 419                 size = 8;
 420                 o(0x50 + vtop->r2); /* push r */
 421             } else {
 422                 size = 4;
 423             }
 424             o(0x50 + r); /* push r */
 425             args_size += size;
 426         }
 427         vtop--;
 428     }
 429     save_regs(0); /* save used temporary registers */
 430     func_sym = vtop->type.ref;
 431     func_call = FUNC_CALL(func_sym->r);
 432     /* fast call case */
 433     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 434         func_call == FUNC_FASTCALLW) {
 435         int fastcall_nb_regs;
 436         uint8_t *fastcall_regs_ptr;
 437         if (func_call == FUNC_FASTCALLW) {
 438             fastcall_regs_ptr = fastcallw_regs;
 439             fastcall_nb_regs = 2;
 440         } else {
 441             fastcall_regs_ptr = fastcall_regs;
 442             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 443         }
 444         for(i = 0;i < fastcall_nb_regs; i++) {
 445             if (args_size <= 0)
 446                 break;
 447             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 448             /* XXX: incorrect for struct/floats */
 449             args_size -= 4;
 450         }
 451     }
 452     gcall_or_jmp(0);
 453
 454 #ifdef TCC_TARGET_PE
 455     if ((func_sym->type.t & VT_BTYPE) == VT_STRUCT)
 456         args_size -= 4;
 457 #endif
 458     if (args_size && func_call != FUNC_STDCALL)
 459         gadd_sp(args_size);
 460     vtop--;
 461 }
 462
 463 #ifdef TCC_TARGET_PE
 464 #define FUNC_PROLOG_SIZE 10
 465 #else
 466 #define FUNC_PROLOG_SIZE 9
 467 #endif
 468
 469 /* generate function prolog of type 't' */
 470 void gfunc_prolog(CType *func_type)
 471 {
 472     int addr, align, size, func_call, fastcall_nb_regs;
 473     int param_index, param_addr;
 474     uint8_t *fastcall_regs_ptr;
 475     Sym *sym;
 476     CType *type;
 477
 478     sym = func_type->ref;
 479     func_call = FUNC_CALL(sym->r);
 480     addr = 8;
 481     loc = 0;
 482     func_vc = 0;
 483
 484     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 485         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 486         fastcall_regs_ptr = fastcall_regs;
 487     } else if (func_call == FUNC_FASTCALLW) {
 488         fastcall_nb_regs = 2;
 489         fastcall_regs_ptr = fastcallw_regs;
 490     } else {
 491         fastcall_nb_regs = 0;
 492         fastcall_regs_ptr = NULL;
 493     }
 494     param_index = 0;
 495
 496     ind += FUNC_PROLOG_SIZE;
 497     func_sub_sp_offset = ind;
 498     /* if the function returns a structure, then add an
 499        implicit pointer parameter */
 500     func_vt = sym->type;
 501     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 502         /* XXX: fastcall case ? */
 503         func_vc = addr;
 504         addr += 4;
 505         param_index++;
 506     }
 507     /* define parameters */
 508     while ((sym = sym->next) != NULL) {
 509         type = &sym->type;
 510         size = type_size(type, &align);
 511         size = (size + 3) & ~3;
 512 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 513         /* structs are passed as pointer */
 514         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 515             size = 4;
 516         }
 517 #endif
 518         if (param_index < fastcall_nb_regs) {
 519             /* save FASTCALL register */
 520             loc -= 4;
 521             o(0x89);     /* movl */
 522             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 523             param_addr = loc;
 524         } else {
 525             param_addr = addr;
 526             addr += size;
 527         }
 528         sym_push(sym->v & ~SYM_FIELD, type,
 529                  VT_LOCAL | lvalue_type(type->t), param_addr);
 530         param_index++;
 531     }
 532     func_ret_sub = 0;
 533     /* pascal type call ? */
 534     if (func_call == FUNC_STDCALL)
 535         func_ret_sub = addr - 8;
 536 #ifdef TCC_TARGET_PE
 537     else if (func_vc)
 538         func_ret_sub = 4;
 539 #endif
 540
 541     /* leave some room for bound checking code */
 542     if (tcc_state->do_bounds_check) {
 543         oad(0xb8, 0); /* lbound section pointer */
 544         oad(0xb8, 0); /* call to function */
 545         func_bound_offset = lbounds_section->data_offset;
 546     }
 547 }
 548
 549 /* generate function epilog */
 550 void gfunc_epilog(void)
 551 {
 552     int v, saved_ind;
 553
 554 #ifdef CONFIG_TCC_BCHECK
 555     if (tcc_state->do_bounds_check
 556      && func_bound_offset != lbounds_section->data_offset) {
 557         int saved_ind;
 558         int *bounds_ptr;
 559         Sym *sym, *sym_data;
 560         /* add end of table info */
 561         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 562         *bounds_ptr = 0;
 563         /* generate bound local allocation */
 564         saved_ind = ind;
 565         ind = func_sub_sp_offset;
 566         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 567                                func_bound_offset, lbounds_section->data_offset);
 568         greloc(cur_text_section, sym_data,
 569                ind + 1, R_386_32);
 570         oad(0xb8, 0); /* mov %eax, xxx */
 571         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 572         greloc(cur_text_section, sym,
 573                ind + 1, R_386_PC32);
 574         oad(0xe8, -4);
 575         ind = saved_ind;
 576         /* generate bound check local freeing */
 577         o(0x5250); /* save returned value, if any */
 578         greloc(cur_text_section, sym_data,
 579                ind + 1, R_386_32);
 580         oad(0xb8, 0); /* mov %eax, xxx */
 581         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 582         greloc(cur_text_section, sym,
 583                ind + 1, R_386_PC32);
 584         oad(0xe8, -4);
 585         o(0x585a); /* restore returned value, if any */
 586     }
 587 #endif
 588     o(0xc9); /* leave */
 589     if (func_ret_sub == 0) {
 590         o(0xc3); /* ret */
 591     } else {
 592         o(0xc2); /* ret n */
 593         g(func_ret_sub);
 594         g(func_ret_sub >> 8);
 595     }
 596     /* align local size to word & save local variables */
 597
 598     v = (-loc + 3) & -4;
 599     saved_ind = ind;
 600     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 601 #ifdef TCC_TARGET_PE
 602     if (v >= 4096) {
 603         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 604         oad(0xb8, v); /* mov stacksize, %eax */
 605         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 606         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 607     } else
 608 #endif
 609     {
 610         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 611         o(0xec81);  /* sub esp, stacksize */
 612         gen_le32(v);
 613 #if FUNC_PROLOG_SIZE == 10
 614         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 615 #endif
 616     }
 617     ind = saved_ind;
 618 }
 619
 620 /* generate a jump to a label */
 621 int gjmp(int t)
 622 {
 623     return psym(0xe9, t);
 624 }
 625
 626 /* generate a jump to a fixed address */
 627 void gjmp_addr(int a)
 628 {
 629     int r;
 630     r = a - ind - 2;
 631     if (r == (char)r) {
 632         g(0xeb);
 633         g(r);
 634     } else {
 635         oad(0xe9, a - ind - 5);
 636     }
 637 }
 638
 639 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 640 int gtst(int inv, int t)
 641 {
 642     int v, *p;
 643
 644     v = vtop->r & VT_VALMASK;
 645     if (v == VT_CMP) {
 646         /* fast case : can jump directly since flags are set */
 647         g(0x0f);
 648         t = psym((vtop->c.i - 16) ^ inv, t);
 649     } else if (v == VT_JMP || v == VT_JMPI) {
 650         /* && or || optimization */
 651         if ((v & 1) == inv) {
 652             /* insert vtop->c jump list in t */
 653             p = &vtop->c.i;
 654             while (*p != 0)
 655                 p = (int *)(cur_text_section->data + *p);
 656             *p = t;
 657             t = vtop->c.i;
 658         } else {
 659             t = gjmp(t);
 660             gsym(vtop->c.i);
 661         }
 662     } else {
 663         if (is_float(vtop->type.t) ||
 664             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 665             vpushi(0);
 666             gen_op(TOK_NE);
 667         }
 668         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 669             /* constant jmp optimization */
 670             if ((vtop->c.i != 0) != inv)
 671                 t = gjmp(t);
 672         } else {
 673             v = gv(RC_INT);
 674             o(0x85);
 675             o(0xc0 + v * 9);
 676             g(0x0f);
 677             t = psym(0x85 ^ inv, t);
 678         }
 679     }
 680     vtop--;
 681     return t;
 682 }
 683
 684 /* generate an integer binary operation */
 685 void gen_opi(int op)
 686 {
 687     int r, fr, opc, c;
 688
 689     switch(op) {
 690     case '+':
 691     case TOK_ADDC1: /* add with carry generation */
 692         opc = 0;
 693     gen_op8:
 694         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 695             /* constant case */
 696             vswap();
 697             r = gv(RC_INT);
 698             vswap();
 699             c = vtop->c.i;
 700             if (c == (char)c) {
 701                 /* XXX: generate inc and dec for smaller code ? */
 702                 o(0x83);
 703                 o(0xc0 | (opc << 3) | r);
 704                 g(c);
 705             } else {
 706                 o(0x81);
 707                 oad(0xc0 | (opc << 3) | r, c);
 708             }
 709         } else {
 710             gv2(RC_INT, RC_INT);
 711             r = vtop[-1].r;
 712             fr = vtop[0].r;
 713             o((opc << 3) | 0x01);
 714             o(0xc0 + r + fr * 8);
 715         }
 716         vtop--;
 717         if (op >= TOK_ULT && op <= TOK_GT) {
 718             vtop->r = VT_CMP;
 719             vtop->c.i = op;
 720         }
 721         break;
 722     case '-':
 723     case TOK_SUBC1: /* sub with carry generation */
 724         opc = 5;
 725         goto gen_op8;
 726     case TOK_ADDC2: /* add with carry use */
 727         opc = 2;
 728         goto gen_op8;
 729     case TOK_SUBC2: /* sub with carry use */
 730         opc = 3;
 731         goto gen_op8;
 732     case '&':
 733         opc = 4;
 734         goto gen_op8;
 735     case '^':
 736         opc = 6;
 737         goto gen_op8;
 738     case '|':
 739         opc = 1;
 740         goto gen_op8;
 741     case '*':
 742         gv2(RC_INT, RC_INT);
 743         r = vtop[-1].r;
 744         fr = vtop[0].r;
 745         vtop--;
 746         o(0xaf0f); /* imul fr, r */
 747         o(0xc0 + fr + r * 8);
 748         break;
 749     case TOK_SHL:
 750         opc = 4;
 751         goto gen_shift;
 752     case TOK_SHR:
 753         opc = 5;
 754         goto gen_shift;
 755     case TOK_SAR:
 756         opc = 7;
 757     gen_shift:
 758         opc = 0xc0 | (opc << 3);
 759         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 760             /* constant case */
 761             vswap();
 762             r = gv(RC_INT);
 763             vswap();
 764             c = vtop->c.i & 0x1f;
 765             o(0xc1); /* shl/shr/sar $xxx, r */
 766             o(opc | r);
 767             g(c);
 768         } else {
 769             /* we generate the shift in ecx */
 770             gv2(RC_INT, RC_ECX);
 771             r = vtop[-1].r;
 772             o(0xd3); /* shl/shr/sar %cl, r */
 773             o(opc | r);
 774         }
 775         vtop--;
 776         break;
 777     case '/':
 778     case TOK_UDIV:
 779     case TOK_PDIV:
 780     case '%':
 781     case TOK_UMOD:
 782     case TOK_UMULL:
 783         /* first operand must be in eax */
 784         /* XXX: need better constraint for second operand */
 785         gv2(RC_EAX, RC_ECX);
 786         r = vtop[-1].r;
 787         fr = vtop[0].r;
 788         vtop--;
 789         save_reg(TREG_EDX);
 790         if (op == TOK_UMULL) {
 791             o(0xf7); /* mul fr */
 792             o(0xe0 + fr);
 793             vtop->r2 = TREG_EDX;
 794             r = TREG_EAX;
 795         } else {
 796             if (op == TOK_UDIV || op == TOK_UMOD) {
 797                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 798                 o(0xf0 + fr);
 799             } else {
 800                 o(0xf799); /* cltd, idiv fr, %eax */
 801                 o(0xf8 + fr);
 802             }
 803             if (op == '%' || op == TOK_UMOD)
 804                 r = TREG_EDX;
 805             else
 806                 r = TREG_EAX;
 807         }
 808         vtop->r = r;
 809         break;
 810     default:
 811         opc = 7;
 812         goto gen_op8;
 813     }
 814 }
 815
 816 /* generate a floating point operation 'v = t1 op t2' instruction. The
 817    two operands are guaranted to have the same floating point type */
 818 /* XXX: need to use ST1 too */
 819 void gen_opf(int op)
 820 {
 821     int a, ft, fc, swapped, r;
 822
 823     /* convert constants to memory references */
 824     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 825         vswap();
 826         gv(RC_FLOAT);
 827         vswap();
 828     }
 829     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 830         gv(RC_FLOAT);
 831
 832     /* must put at least one value in the floating point register */
 833     if ((vtop[-1].r & VT_LVAL) &&
 834         (vtop[0].r & VT_LVAL)) {
 835         vswap();
 836         gv(RC_FLOAT);
 837         vswap();
 838     }
 839     swapped = 0;
 840     /* swap the stack if needed so that t1 is the register and t2 is
 841        the memory reference */
 842     if (vtop[-1].r & VT_LVAL) {
 843         vswap();
 844         swapped = 1;
 845     }
 846     if (op >= TOK_ULT && op <= TOK_GT) {
 847         /* load on stack second operand */
 848         load(TREG_ST0, vtop);
 849         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 850         if (op == TOK_GE || op == TOK_GT)
 851             swapped = !swapped;
 852         else if (op == TOK_EQ || op == TOK_NE)
 853             swapped = 0;
 854         if (swapped)
 855             o(0xc9d9); /* fxch %st(1) */
 856         o(0xe9da); /* fucompp */
 857         o(0xe0df); /* fnstsw %ax */
 858         if (op == TOK_EQ) {
 859             o(0x45e480); /* and $0x45, %ah */
 860             o(0x40fC80); /* cmp $0x40, %ah */
 861         } else if (op == TOK_NE) {
 862             o(0x45e480); /* and $0x45, %ah */
 863             o(0x40f480); /* xor $0x40, %ah */
 864             op = TOK_NE;
 865         } else if (op == TOK_GE || op == TOK_LE) {
 866             o(0x05c4f6); /* test $0x05, %ah */
 867             op = TOK_EQ;
 868         } else {
 869             o(0x45c4f6); /* test $0x45, %ah */
 870             op = TOK_EQ;
 871         }
 872         vtop--;
 873         vtop->r = VT_CMP;
 874         vtop->c.i = op;
 875     } else {
 876         /* no memory reference possible for long double operations */
 877         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 878             load(TREG_ST0, vtop);
 879             swapped = !swapped;
 880         }
 881
 882         switch(op) {
 883         default:
 884         case '+':
 885             a = 0;
 886             break;
 887         case '-':
 888             a = 4;
 889             if (swapped)
 890                 a++;
 891             break;
 892         case '*':
 893             a = 1;
 894             break;
 895         case '/':
 896             a = 6;
 897             if (swapped)
 898                 a++;
 899             break;
 900         }
 901         ft = vtop->type.t;
 902         fc = vtop->c.ul;
 903         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 904             o(0xde); /* fxxxp %st, %st(1) */
 905             o(0xc1 + (a << 3));
 906         } else {
 907             /* if saved lvalue, then we must reload it */
 908             r = vtop->r;
 909             if ((r & VT_VALMASK) == VT_LLOCAL) {
 910                 SValue v1;
 911                 r = get_reg(RC_INT);
 912                 v1.type.t = VT_INT;
 913                 v1.r = VT_LOCAL | VT_LVAL;
 914                 v1.c.ul = fc;
 915                 load(r, &v1);
 916                 fc = 0;
 917             }
 918
 919             if ((ft & VT_BTYPE) == VT_DOUBLE)
 920                 o(0xdc);
 921             else
 922                 o(0xd8);
 923             gen_modrm(a, r, vtop->sym, fc);
 924         }
 925         vtop--;
 926     }
 927 }
 928
 929 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 930    and 'long long' cases. */
 931 void gen_cvt_itof(int t)
 932 {
 933     save_reg(TREG_ST0);
 934     gv(RC_INT);
 935     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 936         /* signed long long to float/double/long double (unsigned case
 937            is handled generically) */
 938         o(0x50 + vtop->r2); /* push r2 */
 939         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 940         o(0x242cdf); /* fildll (%esp) */
 941         o(0x08c483); /* add $8, %esp */
 942     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 943                (VT_INT | VT_UNSIGNED)) {
 944         /* unsigned int to float/double/long double */
 945         o(0x6a); /* push $0 */
 946         g(0x00);
 947         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 948         o(0x242cdf); /* fildll (%esp) */
 949         o(0x08c483); /* add $8, %esp */
 950     } else {
 951         /* int to float/double/long double */
 952         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 953         o(0x2404db); /* fildl (%esp) */
 954         o(0x04c483); /* add $4, %esp */
 955     }
 956     vtop->r = TREG_ST0;
 957 }
 958
 959 /* convert fp to int 't' type */
 960 /* XXX: handle long long case */
 961 void gen_cvt_ftoi(int t)
 962 {
 963     int r, r2, size;
 964     Sym *sym;
 965     CType ushort_type;
 966
 967     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 968
 969     gv(RC_FLOAT);
 970     if (t != VT_INT)
 971         size = 8;
 972     else
 973         size = 4;
 974
 975     o(0x2dd9); /* ldcw xxx */
 976     sym = external_global_sym(TOK___tcc_int_fpu_control,
 977                               &ushort_type, VT_LVAL);
 978     greloc(cur_text_section, sym,
 979            ind, R_386_32);
 980     gen_le32(0);
 981
 982     oad(0xec81, size); /* sub $xxx, %esp */
 983     if (size == 4)
 984         o(0x1cdb); /* fistpl */
 985     else
 986         o(0x3cdf); /* fistpll */
 987     o(0x24);
 988     o(0x2dd9); /* ldcw xxx */
 989     sym = external_global_sym(TOK___tcc_fpu_control,
 990                               &ushort_type, VT_LVAL);
 991     greloc(cur_text_section, sym,
 992            ind, R_386_32);
 993     gen_le32(0);
 994
 995     r = get_reg(RC_INT);
 996     o(0x58 + r); /* pop r */
 997     if (size == 8) {
 998         if (t == VT_LLONG) {
 999             vtop->r = r; /* mark reg as used */
1000             r2 = get_reg(RC_INT);
1001             o(0x58 + r2); /* pop r2 */
1002             vtop->r2 = r2;
1003         } else {
1004             o(0x04c483); /* add $4, %esp */
1005         }
1006     }
1007     vtop->r = r;
1008 }
1009
1010 /* convert from one floating point type to another */
1011 void gen_cvt_ftof(int t)
1012 {
1013     /* all we have to do on i386 is to put the float in a register */
1014     gv(RC_FLOAT);
1015 }
1016
1017 /* computed goto support */
1018 void ggoto(void)
1019 {
1020     gcall_or_jmp(1);
1021     vtop--;
1022 }
1023
1024 /* bound check support functions */
1025 #ifdef CONFIG_TCC_BCHECK
1026
1027 /* generate a bounded pointer addition */
1028 void gen_bounded_ptr_add(void)
1029 {
1030     Sym *sym;
1031
1032     /* prepare fast i386 function call (args in eax and edx) */
1033     gv2(RC_EAX, RC_EDX);
1034     /* save all temporary registers */
1035     vtop -= 2;
1036     save_regs(0);
1037     /* do a fast function call */
1038     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
1039     greloc(cur_text_section, sym,
1040            ind + 1, R_386_PC32);
1041     oad(0xe8, -4);
1042     /* returned pointer is in eax */
1043     vtop++;
1044     vtop->r = TREG_EAX | VT_BOUNDED;
1045     /* address of bounding function call point */
1046     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1047 }
1048
1049 /* patch pointer addition in vtop so that pointer dereferencing is
1050    also tested */
1051 void gen_bounded_ptr_deref(void)
1052 {
1053     int func;
1054     int size, align;
1055     Elf32_Rel *rel;
1056     Sym *sym;
1057
1058     size = 0;
1059     /* XXX: put that code in generic part of tcc */
1060     if (!is_float(vtop->type.t)) {
1061         if (vtop->r & VT_LVAL_BYTE)
1062             size = 1;
1063         else if (vtop->r & VT_LVAL_SHORT)
1064             size = 2;
1065     }
1066     if (!size)
1067         size = type_size(&vtop->type, &align);
1068     switch(size) {
1069     case  1: func = TOK___bound_ptr_indir1; break;
1070     case  2: func = TOK___bound_ptr_indir2; break;
1071     case  4: func = TOK___bound_ptr_indir4; break;
1072     case  8: func = TOK___bound_ptr_indir8; break;
1073     case 12: func = TOK___bound_ptr_indir12; break;
1074     case 16: func = TOK___bound_ptr_indir16; break;
1075     default:
1076         error("unhandled size when derefencing bounded pointer");
1077         func = 0;
1078         break;
1079     }
1080
1081     /* patch relocation */
1082     /* XXX: find a better solution ? */
1083     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1084     sym = external_global_sym(func, &func_old_type, 0);
1085     if (!sym->c)
1086         put_extern_sym(sym, NULL, 0, 0);
1087     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1088 }
1089 #endif
1090
1091 /* end of X86 code generator */
1092 /*************************************************************/
1093