i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001-2004 Fabrice Bellard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  19  */
  20
  21 #ifdef TARGET_DEFS_ONLY
  22
  23 /* number of available registers */
  24 #define NB_REGS         4
  25 #define NB_ASM_REGS     8
  26
  27 /* a register can belong to several classes. The classes must be
  28    sorted from more general to more precise (see gv2() code which does
  29    assumptions on it). */
  30 #define RC_INT     0x0001 /* generic integer register */
  31 #define RC_FLOAT   0x0002 /* generic float register */
  32 #define RC_EAX     0x0004
  33 #define RC_ST0     0x0008
  34 #define RC_ECX     0x0010
  35 #define RC_EDX     0x0020
  36 #define RC_IRET    RC_EAX /* function return: integer register */
  37 #define RC_LRET    RC_EDX /* function return: second integer register */
  38 #define RC_FRET    RC_ST0 /* function return: float register */
  39
  40 /* pretty names for the registers */
  41 enum {
  42     TREG_EAX = 0,
  43     TREG_ECX,
  44     TREG_EDX,
  45     TREG_ST0,
  46 };
  47
  48 /* return registers for function */
  49 #define REG_IRET TREG_EAX /* single word int return register */
  50 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  51 #define REG_FRET TREG_ST0 /* float return register */
  52
  53 /* defined if function parameters must be evaluated in reverse order */
  54 #define INVERT_FUNC_PARAMS
  55
  56 /* defined if structures are passed as pointers. Otherwise structures
  57    are directly pushed on stack. */
  58 //#define FUNC_STRUCT_PARAM_AS_PTR
  59
  60 /* pointer size, in bytes */
  61 #define PTR_SIZE 4
  62
  63 /* long double size and alignment, in bytes */
  64 #define LDOUBLE_SIZE  12
  65 #define LDOUBLE_ALIGN 4
  66 /* maximum alignment (for aligned attribute support) */
  67 #define MAX_ALIGN     8
  68
  69
  70 #define psym oad
  71
  72 /******************************************************/
  73 /* ELF defines */
  74
  75 #define EM_TCC_TARGET EM_386
  76
  77 /* relocation type for 32 bit data relocation */
  78 #define R_DATA_32   R_386_32
  79 #define R_DATA_PTR  R_386_32
  80 #define R_JMP_SLOT  R_386_JMP_SLOT
  81 #define R_COPY      R_386_COPY
  82
  83 #define ELF_START_ADDR 0x08048000
  84 #define ELF_PAGE_SIZE  0x1000
  85
  86 /******************************************************/
  87 #else /* ! TARGET_DEFS_ONLY */
  88 /******************************************************/
  89 #include "tcc.h"
  90
  91 ST_DATA const int reg_classes[NB_REGS] = {
  92     /* eax */ RC_INT | RC_EAX,
  93     /* ecx */ RC_INT | RC_ECX,
  94     /* edx */ RC_INT | RC_EDX,
  95     /* st0 */ RC_FLOAT | RC_ST0,
  96 };
  97
  98 static unsigned long func_sub_sp_offset;
  99 static int func_ret_sub;
 100 #ifdef CONFIG_TCC_BCHECK
 101 static unsigned long func_bound_offset;
 102 #endif
 103
 104 /* XXX: make it faster ? */
 105 ST_FUNC void g(int c)
 106 {
 107     int ind1;
 108     ind1 = ind + 1;
 109     if (ind1 > cur_text_section->data_allocated)
 110         section_realloc(cur_text_section, ind1);
 111     cur_text_section->data[ind] = c;
 112     ind = ind1;
 113 }
 114
 115 ST_FUNC void o(unsigned int c)
 116 {
 117     while (c) {
 118         g(c);
 119         c = c >> 8;
 120     }
 121 }
 122
 123 ST_FUNC void gen_le16(int v)
 124 {
 125     g(v);
 126     g(v >> 8);
 127 }
 128
 129 ST_FUNC void gen_le32(int c)
 130 {
 131     g(c);
 132     g(c >> 8);
 133     g(c >> 16);
 134     g(c >> 24);
 135 }
 136
 137 /* output a symbol and patch all calls to it */
 138 ST_FUNC void gsym_addr(int t, int a)
 139 {
 140     int n, *ptr;
 141     while (t) {
 142         ptr = (int *)(cur_text_section->data + t);
 143         n = *ptr; /* next value */
 144         *ptr = a - t - 4;
 145         t = n;
 146     }
 147 }
 148
 149 ST_FUNC void gsym(int t)
 150 {
 151     gsym_addr(t, ind);
 152 }
 153
 154 /* psym is used to put an instruction with a data field which is a
 155    reference to a symbol. It is in fact the same as oad ! */
 156 #define psym oad
 157
 158 /* instruction + 4 bytes data. Return the address of the data */
 159 ST_FUNC int oad(int c, int s)
 160 {
 161     int ind1;
 162
 163     o(c);
 164     ind1 = ind + 4;
 165     if (ind1 > cur_text_section->data_allocated)
 166         section_realloc(cur_text_section, ind1);
 167     *(int *)(cur_text_section->data + ind) = s;
 168     s = ind;
 169     ind = ind1;
 170     return s;
 171 }
 172
 173 /* output constant with relocation if 'r & VT_SYM' is true */
 174 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 175 {
 176     if (r & VT_SYM)
 177         greloc(cur_text_section, sym, ind, R_386_32);
 178     gen_le32(c);
 179 }
 180
 181 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 182 {
 183     if (r & VT_SYM)
 184         greloc(cur_text_section, sym, ind, R_386_PC32);
 185     gen_le32(c - 4);
 186 }
 187
 188 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 189    opcode bits */
 190 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 191 {
 192     op_reg = op_reg << 3;
 193     if ((r & VT_VALMASK) == VT_CONST) {
 194         /* constant memory reference */
 195         o(0x05 | op_reg);
 196         gen_addr32(r, sym, c);
 197     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 198         /* currently, we use only ebp as base */
 199         if (c == (char)c) {
 200             /* short reference */
 201             o(0x45 | op_reg);
 202             g(c);
 203         } else {
 204             oad(0x85 | op_reg, c);
 205         }
 206     } else {
 207         g(0x00 | op_reg | (r & VT_VALMASK));
 208     }
 209 }
 210
 211 /* load 'r' from value 'sv' */
 212 ST_FUNC void load(int r, SValue *sv)
 213 {
 214     int v, t, ft, fc, fr;
 215     SValue v1;
 216
 217 #ifdef TCC_TARGET_PE
 218     SValue v2;
 219     sv = pe_getimport(sv, &v2);
 220 #endif
 221
 222     fr = sv->r;
 223     ft = sv->type.t;
 224     fc = sv->c.ul;
 225
 226     v = fr & VT_VALMASK;
 227     if (fr & VT_LVAL) {
 228         if (v == VT_LLOCAL) {
 229             v1.type.t = VT_INT;
 230             v1.r = VT_LOCAL | VT_LVAL;
 231             v1.c.ul = fc;
 232             fr = r;
 233             if (!(reg_classes[fr] & RC_INT))
 234                 fr = get_reg(RC_INT);
 235             load(fr, &v1);
 236         }
 237         if ((ft & VT_BTYPE) == VT_FLOAT) {
 238             o(0xd9); /* flds */
 239             r = 0;
 240         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 241             o(0xdd); /* fldl */
 242             r = 0;
 243         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 244             o(0xdb); /* fldt */
 245             r = 5;
 246         } else if ((ft & VT_TYPE) == VT_BYTE) {
 247             o(0xbe0f);   /* movsbl */
 248         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 249             o(0xb60f);   /* movzbl */
 250         } else if ((ft & VT_TYPE) == VT_SHORT) {
 251             o(0xbf0f);   /* movswl */
 252         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 253             o(0xb70f);   /* movzwl */
 254         } else {
 255             o(0x8b);     /* movl */
 256         }
 257         gen_modrm(r, fr, sv->sym, fc);
 258     } else {
 259         if (v == VT_CONST) {
 260             o(0xb8 + r); /* mov $xx, r */
 261             gen_addr32(fr, sv->sym, fc);
 262         } else if (v == VT_LOCAL) {
 263             o(0x8d); /* lea xxx(%ebp), r */
 264             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 265         } else if (v == VT_CMP) {
 266             oad(0xb8 + r, 0); /* mov $0, r */
 267             o(0x0f); /* setxx %br */
 268             o(fc);
 269             o(0xc0 + r);
 270         } else if (v == VT_JMP || v == VT_JMPI) {
 271             t = v & 1;
 272             oad(0xb8 + r, t); /* mov $1, r */
 273             o(0x05eb); /* jmp after */
 274             gsym(fc);
 275             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 276         } else if (v != r) {
 277             o(0x89);
 278             o(0xc0 + r + v * 8); /* mov v, r */
 279         }
 280     }
 281 }
 282
 283 /* store register 'r' in lvalue 'v' */
 284 ST_FUNC void store(int r, SValue *v)
 285 {
 286     int fr, bt, ft, fc;
 287
 288 #ifdef TCC_TARGET_PE
 289     SValue v2;
 290     v = pe_getimport(v, &v2);
 291 #endif
 292
 293     ft = v->type.t;
 294     fc = v->c.ul;
 295     fr = v->r & VT_VALMASK;
 296     bt = ft & VT_BTYPE;
 297     /* XXX: incorrect if float reg to reg */
 298     if (bt == VT_FLOAT) {
 299         o(0xd9); /* fsts */
 300         r = 2;
 301     } else if (bt == VT_DOUBLE) {
 302         o(0xdd); /* fstpl */
 303         r = 2;
 304     } else if (bt == VT_LDOUBLE) {
 305         o(0xc0d9); /* fld %st(0) */
 306         o(0xdb); /* fstpt */
 307         r = 7;
 308     } else {
 309         if (bt == VT_SHORT)
 310             o(0x66);
 311         if (bt == VT_BYTE || bt == VT_BOOL)
 312             o(0x88);
 313         else
 314             o(0x89);
 315     }
 316     if (fr == VT_CONST ||
 317         fr == VT_LOCAL ||
 318         (v->r & VT_LVAL)) {
 319         gen_modrm(r, v->r, v->sym, fc);
 320     } else if (fr != r) {
 321         o(0xc0 + fr + r * 8); /* mov r, fr */
 322     }
 323 }
 324
 325 static void gadd_sp(int val)
 326 {
 327     if (val == (char)val) {
 328         o(0xc483);
 329         g(val);
 330     } else {
 331         oad(0xc481, val); /* add $xxx, %esp */
 332     }
 333 }
 334
 335 /* 'is_jmp' is '1' if it is a jump */
 336 static void gcall_or_jmp(int is_jmp)
 337 {
 338     int r;
 339     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 340         /* constant case */
 341         if (vtop->r & VT_SYM) {
 342             /* relocation case */
 343             greloc(cur_text_section, vtop->sym,
 344                    ind + 1, R_386_PC32);
 345         } else {
 346             /* put an empty PC32 relocation */
 347             put_elf_reloc(symtab_section, cur_text_section,
 348                           ind + 1, R_386_PC32, 0);
 349         }
 350         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 351     } else {
 352         /* otherwise, indirect call */
 353         r = gv(RC_INT);
 354         o(0xff); /* call/jmp *r */
 355         o(0xd0 + r + (is_jmp << 4));
 356     }
 357 }
 358
 359 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 360 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 361
 362 /* Generate function call. The function address is pushed first, then
 363    all the parameters in call order. This functions pops all the
 364    parameters and the function address. */
 365 ST_FUNC void gfunc_call(int nb_args)
 366 {
 367     int size, align, r, args_size, i, func_call;
 368     Sym *func_sym;
 369
 370     args_size = 0;
 371     for(i = 0;i < nb_args; i++) {
 372         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 373             size = type_size(&vtop->type, &align);
 374             /* align to stack align size */
 375             size = (size + 3) & ~3;
 376             /* allocate the necessary size on stack */
 377             oad(0xec81, size); /* sub $xxx, %esp */
 378             /* generate structure store */
 379             r = get_reg(RC_INT);
 380             o(0x89); /* mov %esp, r */
 381             o(0xe0 + r);
 382             vset(&vtop->type, r | VT_LVAL, 0);
 383             vswap();
 384             vstore();
 385             args_size += size;
 386         } else if (is_float(vtop->type.t)) {
 387             gv(RC_FLOAT); /* only one float register */
 388             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 389                 size = 4;
 390             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 391                 size = 8;
 392             else
 393                 size = 12;
 394             oad(0xec81, size); /* sub $xxx, %esp */
 395             if (size == 12)
 396                 o(0x7cdb);
 397             else
 398                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 399             g(0x24);
 400             g(0x00);
 401             args_size += size;
 402         } else {
 403             /* simple type (currently always same size) */
 404             /* XXX: implicit cast ? */
 405             r = gv(RC_INT);
 406             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 407                 size = 8;
 408                 o(0x50 + vtop->r2); /* push r */
 409             } else {
 410                 size = 4;
 411             }
 412             o(0x50 + r); /* push r */
 413             args_size += size;
 414         }
 415         vtop--;
 416     }
 417     save_regs(0); /* save used temporary registers */
 418     func_sym = vtop->type.ref;
 419     func_call = FUNC_CALL(func_sym->r);
 420     /* fast call case */
 421     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 422         func_call == FUNC_FASTCALLW) {
 423         int fastcall_nb_regs;
 424         uint8_t *fastcall_regs_ptr;
 425         if (func_call == FUNC_FASTCALLW) {
 426             fastcall_regs_ptr = fastcallw_regs;
 427             fastcall_nb_regs = 2;
 428         } else {
 429             fastcall_regs_ptr = fastcall_regs;
 430             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 431         }
 432         for(i = 0;i < fastcall_nb_regs; i++) {
 433             if (args_size <= 0)
 434                 break;
 435             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 436             /* XXX: incorrect for struct/floats */
 437             args_size -= 4;
 438         }
 439     }
 440     gcall_or_jmp(0);
 441
 442 #ifdef TCC_TARGET_PE
 443     if ((func_sym->type.t & VT_BTYPE) == VT_STRUCT)
 444         args_size -= 4;
 445 #endif
 446     if (args_size && func_call != FUNC_STDCALL)
 447         gadd_sp(args_size);
 448     vtop--;
 449 }
 450
 451 #ifdef TCC_TARGET_PE
 452 #define FUNC_PROLOG_SIZE 10
 453 #else
 454 #define FUNC_PROLOG_SIZE 9
 455 #endif
 456
 457 /* generate function prolog of type 't' */
 458 ST_FUNC void gfunc_prolog(CType *func_type)
 459 {
 460     int addr, align, size, func_call, fastcall_nb_regs;
 461     int param_index, param_addr;
 462     uint8_t *fastcall_regs_ptr;
 463     Sym *sym;
 464     CType *type;
 465
 466     sym = func_type->ref;
 467     func_call = FUNC_CALL(sym->r);
 468     addr = 8;
 469     loc = 0;
 470     func_vc = 0;
 471
 472     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 473         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 474         fastcall_regs_ptr = fastcall_regs;
 475     } else if (func_call == FUNC_FASTCALLW) {
 476         fastcall_nb_regs = 2;
 477         fastcall_regs_ptr = fastcallw_regs;
 478     } else {
 479         fastcall_nb_regs = 0;
 480         fastcall_regs_ptr = NULL;
 481     }
 482     param_index = 0;
 483
 484     ind += FUNC_PROLOG_SIZE;
 485     func_sub_sp_offset = ind;
 486     /* if the function returns a structure, then add an
 487        implicit pointer parameter */
 488     func_vt = sym->type;
 489     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 490         /* XXX: fastcall case ? */
 491         func_vc = addr;
 492         addr += 4;
 493         param_index++;
 494     }
 495     /* define parameters */
 496     while ((sym = sym->next) != NULL) {
 497         type = &sym->type;
 498         size = type_size(type, &align);
 499         size = (size + 3) & ~3;
 500 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 501         /* structs are passed as pointer */
 502         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 503             size = 4;
 504         }
 505 #endif
 506         if (param_index < fastcall_nb_regs) {
 507             /* save FASTCALL register */
 508             loc -= 4;
 509             o(0x89);     /* movl */
 510             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 511             param_addr = loc;
 512         } else {
 513             param_addr = addr;
 514             addr += size;
 515         }
 516         sym_push(sym->v & ~SYM_FIELD, type,
 517                  VT_LOCAL | lvalue_type(type->t), param_addr);
 518         param_index++;
 519     }
 520     func_ret_sub = 0;
 521     /* pascal type call ? */
 522     if (func_call == FUNC_STDCALL)
 523         func_ret_sub = addr - 8;
 524 #ifdef TCC_TARGET_PE
 525     else if (func_vc)
 526         func_ret_sub = 4;
 527 #endif
 528
 529 #ifdef CONFIG_TCC_BCHECK
 530     /* leave some room for bound checking code */
 531     if (tcc_state->do_bounds_check) {
 532         oad(0xb8, 0); /* lbound section pointer */
 533         oad(0xb8, 0); /* call to function */
 534         func_bound_offset = lbounds_section->data_offset;
 535     }
 536 #endif
 537 }
 538
 539 /* generate function epilog */
 540 ST_FUNC void gfunc_epilog(void)
 541 {
 542     int v, saved_ind;
 543
 544 #ifdef CONFIG_TCC_BCHECK
 545     if (tcc_state->do_bounds_check
 546      && func_bound_offset != lbounds_section->data_offset) {
 547         int saved_ind;
 548         int *bounds_ptr;
 549         Sym *sym, *sym_data;
 550         /* add end of table info */
 551         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 552         *bounds_ptr = 0;
 553         /* generate bound local allocation */
 554         saved_ind = ind;
 555         ind = func_sub_sp_offset;
 556         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 557                                func_bound_offset, lbounds_section->data_offset);
 558         greloc(cur_text_section, sym_data,
 559                ind + 1, R_386_32);
 560         oad(0xb8, 0); /* mov %eax, xxx */
 561         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 562         greloc(cur_text_section, sym,
 563                ind + 1, R_386_PC32);
 564         oad(0xe8, -4);
 565         ind = saved_ind;
 566         /* generate bound check local freeing */
 567         o(0x5250); /* save returned value, if any */
 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_delete, &func_old_type, 0);
 572         greloc(cur_text_section, sym,
 573                ind + 1, R_386_PC32);
 574         oad(0xe8, -4);
 575         o(0x585a); /* restore returned value, if any */
 576     }
 577 #endif
 578     o(0xc9); /* leave */
 579     if (func_ret_sub == 0) {
 580         o(0xc3); /* ret */
 581     } else {
 582         o(0xc2); /* ret n */
 583         g(func_ret_sub);
 584         g(func_ret_sub >> 8);
 585     }
 586     /* align local size to word & save local variables */
 587
 588     v = (-loc + 3) & -4;
 589     saved_ind = ind;
 590     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 591 #ifdef TCC_TARGET_PE
 592     if (v >= 4096) {
 593         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 594         oad(0xb8, v); /* mov stacksize, %eax */
 595         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 596         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 597     } else
 598 #endif
 599     {
 600         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 601         o(0xec81);  /* sub esp, stacksize */
 602         gen_le32(v);
 603 #if FUNC_PROLOG_SIZE == 10
 604         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 605 #endif
 606     }
 607     ind = saved_ind;
 608 }
 609
 610 /* generate a jump to a label */
 611 ST_FUNC int gjmp(int t)
 612 {
 613     return psym(0xe9, t);
 614 }
 615
 616 /* generate a jump to a fixed address */
 617 ST_FUNC void gjmp_addr(int a)
 618 {
 619     int r;
 620     r = a - ind - 2;
 621     if (r == (char)r) {
 622         g(0xeb);
 623         g(r);
 624     } else {
 625         oad(0xe9, a - ind - 5);
 626     }
 627 }
 628
 629 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 630 ST_FUNC int gtst(int inv, int t)
 631 {
 632     int v, *p;
 633
 634     v = vtop->r & VT_VALMASK;
 635     if (v == VT_CMP) {
 636         /* fast case : can jump directly since flags are set */
 637         g(0x0f);
 638         t = psym((vtop->c.i - 16) ^ inv, t);
 639     } else if (v == VT_JMP || v == VT_JMPI) {
 640         /* && or || optimization */
 641         if ((v & 1) == inv) {
 642             /* insert vtop->c jump list in t */
 643             p = &vtop->c.i;
 644             while (*p != 0)
 645                 p = (int *)(cur_text_section->data + *p);
 646             *p = t;
 647             t = vtop->c.i;
 648         } else {
 649             t = gjmp(t);
 650             gsym(vtop->c.i);
 651         }
 652     } else {
 653         if (is_float(vtop->type.t) ||
 654             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 655             vpushi(0);
 656             gen_op(TOK_NE);
 657         }
 658         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 659             /* constant jmp optimization */
 660             if ((vtop->c.i != 0) != inv)
 661                 t = gjmp(t);
 662         } else {
 663             v = gv(RC_INT);
 664             o(0x85);
 665             o(0xc0 + v * 9);
 666             g(0x0f);
 667             t = psym(0x85 ^ inv, t);
 668         }
 669     }
 670     vtop--;
 671     return t;
 672 }
 673
 674 /* generate an integer binary operation */
 675 ST_FUNC void gen_opi(int op)
 676 {
 677     int r, fr, opc, c;
 678
 679     switch(op) {
 680     case '+':
 681     case TOK_ADDC1: /* add with carry generation */
 682         opc = 0;
 683     gen_op8:
 684         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 685             /* constant case */
 686             vswap();
 687             r = gv(RC_INT);
 688             vswap();
 689             c = vtop->c.i;
 690             if (c == (char)c) {
 691                 /* generate inc and dec for smaller code */
 692                 if (c==1 && opc==0) {
 693                     o (0x40 | r); // inc
 694                 } else if (c==1 && opc==5) {
 695                     o (0x48 | r); // dec
 696                 } else {
 697                     o(0x83);
 698                     o(0xc0 | (opc << 3) | r);
 699                     g(c);
 700                 }
 701             } else {
 702                 o(0x81);
 703                 oad(0xc0 | (opc << 3) | r, c);
 704             }
 705         } else {
 706             gv2(RC_INT, RC_INT);
 707             r = vtop[-1].r;
 708             fr = vtop[0].r;
 709             o((opc << 3) | 0x01);
 710             o(0xc0 + r + fr * 8);
 711         }
 712         vtop--;
 713         if (op >= TOK_ULT && op <= TOK_GT) {
 714             vtop->r = VT_CMP;
 715             vtop->c.i = op;
 716         }
 717         break;
 718     case '-':
 719     case TOK_SUBC1: /* sub with carry generation */
 720         opc = 5;
 721         goto gen_op8;
 722     case TOK_ADDC2: /* add with carry use */
 723         opc = 2;
 724         goto gen_op8;
 725     case TOK_SUBC2: /* sub with carry use */
 726         opc = 3;
 727         goto gen_op8;
 728     case '&':
 729         opc = 4;
 730         goto gen_op8;
 731     case '^':
 732         opc = 6;
 733         goto gen_op8;
 734     case '|':
 735         opc = 1;
 736         goto gen_op8;
 737     case '*':
 738         gv2(RC_INT, RC_INT);
 739         r = vtop[-1].r;
 740         fr = vtop[0].r;
 741         vtop--;
 742         o(0xaf0f); /* imul fr, r */
 743         o(0xc0 + fr + r * 8);
 744         break;
 745     case TOK_SHL:
 746         opc = 4;
 747         goto gen_shift;
 748     case TOK_SHR:
 749         opc = 5;
 750         goto gen_shift;
 751     case TOK_SAR:
 752         opc = 7;
 753     gen_shift:
 754         opc = 0xc0 | (opc << 3);
 755         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 756             /* constant case */
 757             vswap();
 758             r = gv(RC_INT);
 759             vswap();
 760             c = vtop->c.i & 0x1f;
 761             o(0xc1); /* shl/shr/sar $xxx, r */
 762             o(opc | r);
 763             g(c);
 764         } else {
 765             /* we generate the shift in ecx */
 766             gv2(RC_INT, RC_ECX);
 767             r = vtop[-1].r;
 768             o(0xd3); /* shl/shr/sar %cl, r */
 769             o(opc | r);
 770         }
 771         vtop--;
 772         break;
 773     case '/':
 774     case TOK_UDIV:
 775     case TOK_PDIV:
 776     case '%':
 777     case TOK_UMOD:
 778     case TOK_UMULL:
 779         /* first operand must be in eax */
 780         /* XXX: need better constraint for second operand */
 781         gv2(RC_EAX, RC_ECX);
 782         r = vtop[-1].r;
 783         fr = vtop[0].r;
 784         vtop--;
 785         save_reg(TREG_EDX);
 786         if (op == TOK_UMULL) {
 787             o(0xf7); /* mul fr */
 788             o(0xe0 + fr);
 789             vtop->r2 = TREG_EDX;
 790             r = TREG_EAX;
 791         } else {
 792             if (op == TOK_UDIV || op == TOK_UMOD) {
 793                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 794                 o(0xf0 + fr);
 795             } else {
 796                 o(0xf799); /* cltd, idiv fr, %eax */
 797                 o(0xf8 + fr);
 798             }
 799             if (op == '%' || op == TOK_UMOD)
 800                 r = TREG_EDX;
 801             else
 802                 r = TREG_EAX;
 803         }
 804         vtop->r = r;
 805         break;
 806     default:
 807         opc = 7;
 808         goto gen_op8;
 809     }
 810 }
 811
 812 /* generate a floating point operation 'v = t1 op t2' instruction. The
 813    two operands are guaranted to have the same floating point type */
 814 /* XXX: need to use ST1 too */
 815 ST_FUNC void gen_opf(int op)
 816 {
 817     int a, ft, fc, swapped, r;
 818
 819     /* convert constants to memory references */
 820     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 821         vswap();
 822         gv(RC_FLOAT);
 823         vswap();
 824     }
 825     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 826         gv(RC_FLOAT);
 827
 828     /* must put at least one value in the floating point register */
 829     if ((vtop[-1].r & VT_LVAL) &&
 830         (vtop[0].r & VT_LVAL)) {
 831         vswap();
 832         gv(RC_FLOAT);
 833         vswap();
 834     }
 835     swapped = 0;
 836     /* swap the stack if needed so that t1 is the register and t2 is
 837        the memory reference */
 838     if (vtop[-1].r & VT_LVAL) {
 839         vswap();
 840         swapped = 1;
 841     }
 842     if (op >= TOK_ULT && op <= TOK_GT) {
 843         /* load on stack second operand */
 844         load(TREG_ST0, vtop);
 845         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 846         if (op == TOK_GE || op == TOK_GT)
 847             swapped = !swapped;
 848         else if (op == TOK_EQ || op == TOK_NE)
 849             swapped = 0;
 850         if (swapped)
 851             o(0xc9d9); /* fxch %st(1) */
 852         o(0xe9da); /* fucompp */
 853         o(0xe0df); /* fnstsw %ax */
 854         if (op == TOK_EQ) {
 855             o(0x45e480); /* and $0x45, %ah */
 856             o(0x40fC80); /* cmp $0x40, %ah */
 857         } else if (op == TOK_NE) {
 858             o(0x45e480); /* and $0x45, %ah */
 859             o(0x40f480); /* xor $0x40, %ah */
 860             op = TOK_NE;
 861         } else if (op == TOK_GE || op == TOK_LE) {
 862             o(0x05c4f6); /* test $0x05, %ah */
 863             op = TOK_EQ;
 864         } else {
 865             o(0x45c4f6); /* test $0x45, %ah */
 866             op = TOK_EQ;
 867         }
 868         vtop--;
 869         vtop->r = VT_CMP;
 870         vtop->c.i = op;
 871     } else {
 872         /* no memory reference possible for long double operations */
 873         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 874             load(TREG_ST0, vtop);
 875             swapped = !swapped;
 876         }
 877
 878         switch(op) {
 879         default:
 880         case '+':
 881             a = 0;
 882             break;
 883         case '-':
 884             a = 4;
 885             if (swapped)
 886                 a++;
 887             break;
 888         case '*':
 889             a = 1;
 890             break;
 891         case '/':
 892             a = 6;
 893             if (swapped)
 894                 a++;
 895             break;
 896         }
 897         ft = vtop->type.t;
 898         fc = vtop->c.ul;
 899         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 900             o(0xde); /* fxxxp %st, %st(1) */
 901             o(0xc1 + (a << 3));
 902         } else {
 903             /* if saved lvalue, then we must reload it */
 904             r = vtop->r;
 905             if ((r & VT_VALMASK) == VT_LLOCAL) {
 906                 SValue v1;
 907                 r = get_reg(RC_INT);
 908                 v1.type.t = VT_INT;
 909                 v1.r = VT_LOCAL | VT_LVAL;
 910                 v1.c.ul = fc;
 911                 load(r, &v1);
 912                 fc = 0;
 913             }
 914
 915             if ((ft & VT_BTYPE) == VT_DOUBLE)
 916                 o(0xdc);
 917             else
 918                 o(0xd8);
 919             gen_modrm(a, r, vtop->sym, fc);
 920         }
 921         vtop--;
 922     }
 923 }
 924
 925 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 926    and 'long long' cases. */
 927 ST_FUNC void gen_cvt_itof(int t)
 928 {
 929     save_reg(TREG_ST0);
 930     gv(RC_INT);
 931     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 932         /* signed long long to float/double/long double (unsigned case
 933            is handled generically) */
 934         o(0x50 + vtop->r2); /* push r2 */
 935         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 936         o(0x242cdf); /* fildll (%esp) */
 937         o(0x08c483); /* add $8, %esp */
 938     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 939                (VT_INT | VT_UNSIGNED)) {
 940         /* unsigned int to float/double/long double */
 941         o(0x6a); /* push $0 */
 942         g(0x00);
 943         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 944         o(0x242cdf); /* fildll (%esp) */
 945         o(0x08c483); /* add $8, %esp */
 946     } else {
 947         /* int to float/double/long double */
 948         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 949         o(0x2404db); /* fildl (%esp) */
 950         o(0x04c483); /* add $4, %esp */
 951     }
 952     vtop->r = TREG_ST0;
 953 }
 954
 955 /* convert fp to int 't' type */
 956 /* XXX: handle long long case */
 957 ST_FUNC void gen_cvt_ftoi(int t)
 958 {
 959     int r, r2, size;
 960     Sym *sym;
 961     CType ushort_type;
 962
 963     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 964     ushort_type.ref = 0;
 965
 966     gv(RC_FLOAT);
 967     if (t != VT_INT)
 968         size = 8;
 969     else
 970         size = 4;
 971
 972     o(0x2dd9); /* ldcw xxx */
 973     sym = external_global_sym(TOK___tcc_int_fpu_control,
 974                               &ushort_type, VT_LVAL);
 975     greloc(cur_text_section, sym,
 976            ind, R_386_32);
 977     gen_le32(0);
 978
 979     oad(0xec81, size); /* sub $xxx, %esp */
 980     if (size == 4)
 981         o(0x1cdb); /* fistpl */
 982     else
 983         o(0x3cdf); /* fistpll */
 984     o(0x24);
 985     o(0x2dd9); /* ldcw xxx */
 986     sym = external_global_sym(TOK___tcc_fpu_control,
 987                               &ushort_type, VT_LVAL);
 988     greloc(cur_text_section, sym,
 989            ind, R_386_32);
 990     gen_le32(0);
 991
 992     r = get_reg(RC_INT);
 993     o(0x58 + r); /* pop r */
 994     if (size == 8) {
 995         if (t == VT_LLONG) {
 996             vtop->r = r; /* mark reg as used */
 997             r2 = get_reg(RC_INT);
 998             o(0x58 + r2); /* pop r2 */
 999             vtop->r2 = r2;
1000         } else {
1001             o(0x04c483); /* add $4, %esp */
1002         }
1003     }
1004     vtop->r = r;
1005 }
1006
1007 /* convert from one floating point type to another */
1008 ST_FUNC void gen_cvt_ftof(int t)
1009 {
1010     /* all we have to do on i386 is to put the float in a register */
1011     gv(RC_FLOAT);
1012 }
1013
1014 /* computed goto support */
1015 ST_FUNC void ggoto(void)
1016 {
1017     gcall_or_jmp(1);
1018     vtop--;
1019 }
1020
1021 /* bound check support functions */
1022 #ifdef CONFIG_TCC_BCHECK
1023
1024 /* generate a bounded pointer addition */
1025 ST_FUNC void gen_bounded_ptr_add(void)
1026 {
1027     Sym *sym;
1028
1029     /* prepare fast i386 function call (args in eax and edx) */
1030     gv2(RC_EAX, RC_EDX);
1031     /* save all temporary registers */
1032     vtop -= 2;
1033     save_regs(0);
1034     /* do a fast function call */
1035     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
1036     greloc(cur_text_section, sym,
1037            ind + 1, R_386_PC32);
1038     oad(0xe8, -4);
1039     /* returned pointer is in eax */
1040     vtop++;
1041     vtop->r = TREG_EAX | VT_BOUNDED;
1042     /* address of bounding function call point */
1043     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1044 }
1045
1046 /* patch pointer addition in vtop so that pointer dereferencing is
1047    also tested */
1048 ST_FUNC void gen_bounded_ptr_deref(void)
1049 {
1050     int func;
1051     int size, align;
1052     Elf32_Rel *rel;
1053     Sym *sym;
1054
1055     size = 0;
1056     /* XXX: put that code in generic part of tcc */
1057     if (!is_float(vtop->type.t)) {
1058         if (vtop->r & VT_LVAL_BYTE)
1059             size = 1;
1060         else if (vtop->r & VT_LVAL_SHORT)
1061             size = 2;
1062     }
1063     if (!size)
1064         size = type_size(&vtop->type, &align);
1065     switch(size) {
1066     case  1: func = TOK___bound_ptr_indir1; break;
1067     case  2: func = TOK___bound_ptr_indir2; break;
1068     case  4: func = TOK___bound_ptr_indir4; break;
1069     case  8: func = TOK___bound_ptr_indir8; break;
1070     case 12: func = TOK___bound_ptr_indir12; break;
1071     case 16: func = TOK___bound_ptr_indir16; break;
1072     default:
1073         tcc_error("unhandled size when dereferencing bounded pointer");
1074         func = 0;
1075         break;
1076     }
1077
1078     /* patch relocation */
1079     /* XXX: find a better solution ? */
1080     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1081     sym = external_global_sym(func, &func_old_type, 0);
1082     if (!sym->c)
1083         put_extern_sym(sym, NULL, 0, 0);
1084     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1085 }
1086 #endif
1087
1088 /* end of X86 code generator */
1089 /*************************************************************/
1090 #endif
1091 /*************************************************************/