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             load(r, &v1);
 233             fr = r;
 234         }
 235         if ((ft & VT_BTYPE) == VT_FLOAT) {
 236             o(0xd9); /* flds */
 237             r = 0;
 238         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 239             o(0xdd); /* fldl */
 240             r = 0;
 241         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 242             o(0xdb); /* fldt */
 243             r = 5;
 244         } else if ((ft & VT_TYPE) == VT_BYTE) {
 245             o(0xbe0f);   /* movsbl */
 246         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 247             o(0xb60f);   /* movzbl */
 248         } else if ((ft & VT_TYPE) == VT_SHORT) {
 249             o(0xbf0f);   /* movswl */
 250         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 251             o(0xb70f);   /* movzwl */
 252         } else {
 253             o(0x8b);     /* movl */
 254         }
 255         gen_modrm(r, fr, sv->sym, fc);
 256     } else {
 257         if (v == VT_CONST) {
 258             o(0xb8 + r); /* mov $xx, r */
 259             gen_addr32(fr, sv->sym, fc);
 260         } else if (v == VT_LOCAL) {
 261             o(0x8d); /* lea xxx(%ebp), r */
 262             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 263         } else if (v == VT_CMP) {
 264             oad(0xb8 + r, 0); /* mov $0, r */
 265             o(0x0f); /* setxx %br */
 266             o(fc);
 267             o(0xc0 + r);
 268         } else if (v == VT_JMP || v == VT_JMPI) {
 269             t = v & 1;
 270             oad(0xb8 + r, t); /* mov $1, r */
 271             o(0x05eb); /* jmp after */
 272             gsym(fc);
 273             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 274         } else if (v != r) {
 275             o(0x89);
 276             o(0xc0 + r + v * 8); /* mov v, r */
 277         }
 278     }
 279 }
 280
 281 /* store register 'r' in lvalue 'v' */
 282 ST_FUNC void store(int r, SValue *v)
 283 {
 284     int fr, bt, ft, fc;
 285
 286 #ifdef TCC_TARGET_PE
 287     SValue v2;
 288     v = pe_getimport(v, &v2);
 289 #endif
 290
 291     ft = v->type.t;
 292     fc = v->c.ul;
 293     fr = v->r & VT_VALMASK;
 294     bt = ft & VT_BTYPE;
 295     /* XXX: incorrect if float reg to reg */
 296     if (bt == VT_FLOAT) {
 297         o(0xd9); /* fsts */
 298         r = 2;
 299     } else if (bt == VT_DOUBLE) {
 300         o(0xdd); /* fstpl */
 301         r = 2;
 302     } else if (bt == VT_LDOUBLE) {
 303         o(0xc0d9); /* fld %st(0) */
 304         o(0xdb); /* fstpt */
 305         r = 7;
 306     } else {
 307         if (bt == VT_SHORT)
 308             o(0x66);
 309         if (bt == VT_BYTE || bt == VT_BOOL)
 310             o(0x88);
 311         else
 312             o(0x89);
 313     }
 314     if (fr == VT_CONST ||
 315         fr == VT_LOCAL ||
 316         (v->r & VT_LVAL)) {
 317         gen_modrm(r, v->r, v->sym, fc);
 318     } else if (fr != r) {
 319         o(0xc0 + fr + r * 8); /* mov r, fr */
 320     }
 321 }
 322
 323 static void gadd_sp(int val)
 324 {
 325     if (val == (char)val) {
 326         o(0xc483);
 327         g(val);
 328     } else {
 329         oad(0xc481, val); /* add $xxx, %esp */
 330     }
 331 }
 332
 333 /* 'is_jmp' is '1' if it is a jump */
 334 static void gcall_or_jmp(int is_jmp)
 335 {
 336     int r;
 337     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 338         /* constant case */
 339         if (vtop->r & VT_SYM) {
 340             /* relocation case */
 341             greloc(cur_text_section, vtop->sym,
 342                    ind + 1, R_386_PC32);
 343         } else {
 344             /* put an empty PC32 relocation */
 345             put_elf_reloc(symtab_section, cur_text_section,
 346                           ind + 1, R_386_PC32, 0);
 347         }
 348         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 349     } else {
 350         /* otherwise, indirect call */
 351         r = gv(RC_INT);
 352         o(0xff); /* call/jmp *r */
 353         o(0xd0 + r + (is_jmp << 4));
 354     }
 355 }
 356
 357 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 358 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 359
 360 /* Generate function call. The function address is pushed first, then
 361    all the parameters in call order. This functions pops all the
 362    parameters and the function address. */
 363 ST_FUNC void gfunc_call(int nb_args)
 364 {
 365     int size, align, r, args_size, i, func_call;
 366     Sym *func_sym;
 367
 368     args_size = 0;
 369     for(i = 0;i < nb_args; i++) {
 370         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 371             size = type_size(&vtop->type, &align);
 372             /* align to stack align size */
 373             size = (size + 3) & ~3;
 374             /* allocate the necessary size on stack */
 375             oad(0xec81, size); /* sub $xxx, %esp */
 376             /* generate structure store */
 377             r = get_reg(RC_INT);
 378             o(0x89); /* mov %esp, r */
 379             o(0xe0 + r);
 380             vset(&vtop->type, r | VT_LVAL, 0);
 381             vswap();
 382             vstore();
 383             args_size += size;
 384         } else if (is_float(vtop->type.t)) {
 385             gv(RC_FLOAT); /* only one float register */
 386             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 387                 size = 4;
 388             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 389                 size = 8;
 390             else
 391                 size = 12;
 392             oad(0xec81, size); /* sub $xxx, %esp */
 393             if (size == 12)
 394                 o(0x7cdb);
 395             else
 396                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 397             g(0x24);
 398             g(0x00);
 399             args_size += size;
 400         } else {
 401             /* simple type (currently always same size) */
 402             /* XXX: implicit cast ? */
 403             r = gv(RC_INT);
 404             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 405                 size = 8;
 406                 o(0x50 + vtop->r2); /* push r */
 407             } else {
 408                 size = 4;
 409             }
 410             o(0x50 + r); /* push r */
 411             args_size += size;
 412         }
 413         vtop--;
 414     }
 415     save_regs(0); /* save used temporary registers */
 416     func_sym = vtop->type.ref;
 417     func_call = FUNC_CALL(func_sym->r);
 418     /* fast call case */
 419     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 420         func_call == FUNC_FASTCALLW) {
 421         int fastcall_nb_regs;
 422         uint8_t *fastcall_regs_ptr;
 423         if (func_call == FUNC_FASTCALLW) {
 424             fastcall_regs_ptr = fastcallw_regs;
 425             fastcall_nb_regs = 2;
 426         } else {
 427             fastcall_regs_ptr = fastcall_regs;
 428             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 429         }
 430         for(i = 0;i < fastcall_nb_regs; i++) {
 431             if (args_size <= 0)
 432                 break;
 433             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 434             /* XXX: incorrect for struct/floats */
 435             args_size -= 4;
 436         }
 437     }
 438     gcall_or_jmp(0);
 439
 440 #ifdef TCC_TARGET_PE
 441     if ((func_sym->type.t & VT_BTYPE) == VT_STRUCT)
 442         args_size -= 4;
 443 #endif
 444     if (args_size && func_call != FUNC_STDCALL)
 445         gadd_sp(args_size);
 446     vtop--;
 447 }
 448
 449 #ifdef TCC_TARGET_PE
 450 #define FUNC_PROLOG_SIZE 10
 451 #else
 452 #define FUNC_PROLOG_SIZE 9
 453 #endif
 454
 455 /* generate function prolog of type 't' */
 456 ST_FUNC void gfunc_prolog(CType *func_type)
 457 {
 458     int addr, align, size, func_call, fastcall_nb_regs;
 459     int param_index, param_addr;
 460     uint8_t *fastcall_regs_ptr;
 461     Sym *sym;
 462     CType *type;
 463
 464     sym = func_type->ref;
 465     func_call = FUNC_CALL(sym->r);
 466     addr = 8;
 467     loc = 0;
 468     func_vc = 0;
 469
 470     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 471         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 472         fastcall_regs_ptr = fastcall_regs;
 473     } else if (func_call == FUNC_FASTCALLW) {
 474         fastcall_nb_regs = 2;
 475         fastcall_regs_ptr = fastcallw_regs;
 476     } else {
 477         fastcall_nb_regs = 0;
 478         fastcall_regs_ptr = NULL;
 479     }
 480     param_index = 0;
 481
 482     ind += FUNC_PROLOG_SIZE;
 483     func_sub_sp_offset = ind;
 484     /* if the function returns a structure, then add an
 485        implicit pointer parameter */
 486     func_vt = sym->type;
 487     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 488         /* XXX: fastcall case ? */
 489         func_vc = addr;
 490         addr += 4;
 491         param_index++;
 492     }
 493     /* define parameters */
 494     while ((sym = sym->next) != NULL) {
 495         type = &sym->type;
 496         size = type_size(type, &align);
 497         size = (size + 3) & ~3;
 498 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 499         /* structs are passed as pointer */
 500         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 501             size = 4;
 502         }
 503 #endif
 504         if (param_index < fastcall_nb_regs) {
 505             /* save FASTCALL register */
 506             loc -= 4;
 507             o(0x89);     /* movl */
 508             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 509             param_addr = loc;
 510         } else {
 511             param_addr = addr;
 512             addr += size;
 513         }
 514         sym_push(sym->v & ~SYM_FIELD, type,
 515                  VT_LOCAL | lvalue_type(type->t), param_addr);
 516         param_index++;
 517     }
 518     func_ret_sub = 0;
 519     /* pascal type call ? */
 520     if (func_call == FUNC_STDCALL)
 521         func_ret_sub = addr - 8;
 522 #ifdef TCC_TARGET_PE
 523     else if (func_vc)
 524         func_ret_sub = 4;
 525 #endif
 526
 527 #ifdef CONFIG_TCC_BCHECK
 528     /* leave some room for bound checking code */
 529     if (tcc_state->do_bounds_check) {
 530         oad(0xb8, 0); /* lbound section pointer */
 531         oad(0xb8, 0); /* call to function */
 532         func_bound_offset = lbounds_section->data_offset;
 533     }
 534 #endif
 535 }
 536
 537 /* generate function epilog */
 538 ST_FUNC void gfunc_epilog(void)
 539 {
 540     int v, saved_ind;
 541
 542 #ifdef CONFIG_TCC_BCHECK
 543     if (tcc_state->do_bounds_check
 544      && func_bound_offset != lbounds_section->data_offset) {
 545         int saved_ind;
 546         int *bounds_ptr;
 547         Sym *sym, *sym_data;
 548         /* add end of table info */
 549         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 550         *bounds_ptr = 0;
 551         /* generate bound local allocation */
 552         saved_ind = ind;
 553         ind = func_sub_sp_offset;
 554         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 555                                func_bound_offset, lbounds_section->data_offset);
 556         greloc(cur_text_section, sym_data,
 557                ind + 1, R_386_32);
 558         oad(0xb8, 0); /* mov %eax, xxx */
 559         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 560         greloc(cur_text_section, sym,
 561                ind + 1, R_386_PC32);
 562         oad(0xe8, -4);
 563         ind = saved_ind;
 564         /* generate bound check local freeing */
 565         o(0x5250); /* save returned value, if any */
 566         greloc(cur_text_section, sym_data,
 567                ind + 1, R_386_32);
 568         oad(0xb8, 0); /* mov %eax, xxx */
 569         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 570         greloc(cur_text_section, sym,
 571                ind + 1, R_386_PC32);
 572         oad(0xe8, -4);
 573         o(0x585a); /* restore returned value, if any */
 574     }
 575 #endif
 576     o(0xc9); /* leave */
 577     if (func_ret_sub == 0) {
 578         o(0xc3); /* ret */
 579     } else {
 580         o(0xc2); /* ret n */
 581         g(func_ret_sub);
 582         g(func_ret_sub >> 8);
 583     }
 584     /* align local size to word & save local variables */
 585
 586     v = (-loc + 3) & -4;
 587     saved_ind = ind;
 588     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 589 #ifdef TCC_TARGET_PE
 590     if (v >= 4096) {
 591         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 592         oad(0xb8, v); /* mov stacksize, %eax */
 593         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 594         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 595     } else
 596 #endif
 597     {
 598         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 599         o(0xec81);  /* sub esp, stacksize */
 600         gen_le32(v);
 601 #if FUNC_PROLOG_SIZE == 10
 602         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 603 #endif
 604     }
 605     ind = saved_ind;
 606 }
 607
 608 /* generate a jump to a label */
 609 ST_FUNC int gjmp(int t)
 610 {
 611     return psym(0xe9, t);
 612 }
 613
 614 /* generate a jump to a fixed address */
 615 ST_FUNC void gjmp_addr(int a)
 616 {
 617     int r;
 618     r = a - ind - 2;
 619     if (r == (char)r) {
 620         g(0xeb);
 621         g(r);
 622     } else {
 623         oad(0xe9, a - ind - 5);
 624     }
 625 }
 626
 627 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 628 ST_FUNC int gtst(int inv, int t)
 629 {
 630     int v, *p;
 631
 632     v = vtop->r & VT_VALMASK;
 633     if (v == VT_CMP) {
 634         /* fast case : can jump directly since flags are set */
 635         g(0x0f);
 636         t = psym((vtop->c.i - 16) ^ inv, t);
 637     } else if (v == VT_JMP || v == VT_JMPI) {
 638         /* && or || optimization */
 639         if ((v & 1) == inv) {
 640             /* insert vtop->c jump list in t */
 641             p = &vtop->c.i;
 642             while (*p != 0)
 643                 p = (int *)(cur_text_section->data + *p);
 644             *p = t;
 645             t = vtop->c.i;
 646         } else {
 647             t = gjmp(t);
 648             gsym(vtop->c.i);
 649         }
 650     } else {
 651         if (is_float(vtop->type.t) ||
 652             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 653             vpushi(0);
 654             gen_op(TOK_NE);
 655         }
 656         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 657             /* constant jmp optimization */
 658             if ((vtop->c.i != 0) != inv)
 659                 t = gjmp(t);
 660         } else {
 661             v = gv(RC_INT);
 662             o(0x85);
 663             o(0xc0 + v * 9);
 664             g(0x0f);
 665             t = psym(0x85 ^ inv, t);
 666         }
 667     }
 668     vtop--;
 669     return t;
 670 }
 671
 672 /* generate an integer binary operation */
 673 ST_FUNC void gen_opi(int op)
 674 {
 675     int r, fr, opc, c;
 676
 677     switch(op) {
 678     case '+':
 679     case TOK_ADDC1: /* add with carry generation */
 680         opc = 0;
 681     gen_op8:
 682         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 683             /* constant case */
 684             vswap();
 685             r = gv(RC_INT);
 686             vswap();
 687             c = vtop->c.i;
 688             if (c == (char)c) {
 689                 /* generate inc and dec for smaller code */
 690                 if (c==1 && opc==0) {
 691                     o (0x40 | r); // inc
 692                 } else if (c==1 && opc==5) {
 693                     o (0x48 | r); // dec
 694                 } else {
 695                     o(0x83);
 696                     o(0xc0 | (opc << 3) | r);
 697                     g(c);
 698                 }
 699             } else {
 700                 o(0x81);
 701                 oad(0xc0 | (opc << 3) | r, c);
 702             }
 703         } else {
 704             gv2(RC_INT, RC_INT);
 705             r = vtop[-1].r;
 706             fr = vtop[0].r;
 707             o((opc << 3) | 0x01);
 708             o(0xc0 + r + fr * 8);
 709         }
 710         vtop--;
 711         if (op >= TOK_ULT && op <= TOK_GT) {
 712             vtop->r = VT_CMP;
 713             vtop->c.i = op;
 714         }
 715         break;
 716     case '-':
 717     case TOK_SUBC1: /* sub with carry generation */
 718         opc = 5;
 719         goto gen_op8;
 720     case TOK_ADDC2: /* add with carry use */
 721         opc = 2;
 722         goto gen_op8;
 723     case TOK_SUBC2: /* sub with carry use */
 724         opc = 3;
 725         goto gen_op8;
 726     case '&':
 727         opc = 4;
 728         goto gen_op8;
 729     case '^':
 730         opc = 6;
 731         goto gen_op8;
 732     case '|':
 733         opc = 1;
 734         goto gen_op8;
 735     case '*':
 736         gv2(RC_INT, RC_INT);
 737         r = vtop[-1].r;
 738         fr = vtop[0].r;
 739         vtop--;
 740         o(0xaf0f); /* imul fr, r */
 741         o(0xc0 + fr + r * 8);
 742         break;
 743     case TOK_SHL:
 744         opc = 4;
 745         goto gen_shift;
 746     case TOK_SHR:
 747         opc = 5;
 748         goto gen_shift;
 749     case TOK_SAR:
 750         opc = 7;
 751     gen_shift:
 752         opc = 0xc0 | (opc << 3);
 753         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 754             /* constant case */
 755             vswap();
 756             r = gv(RC_INT);
 757             vswap();
 758             c = vtop->c.i & 0x1f;
 759             o(0xc1); /* shl/shr/sar $xxx, r */
 760             o(opc | r);
 761             g(c);
 762         } else {
 763             /* we generate the shift in ecx */
 764             gv2(RC_INT, RC_ECX);
 765             r = vtop[-1].r;
 766             o(0xd3); /* shl/shr/sar %cl, r */
 767             o(opc | r);
 768         }
 769         vtop--;
 770         break;
 771     case '/':
 772     case TOK_UDIV:
 773     case TOK_PDIV:
 774     case '%':
 775     case TOK_UMOD:
 776     case TOK_UMULL:
 777         /* first operand must be in eax */
 778         /* XXX: need better constraint for second operand */
 779         gv2(RC_EAX, RC_ECX);
 780         r = vtop[-1].r;
 781         fr = vtop[0].r;
 782         vtop--;
 783         save_reg(TREG_EDX);
 784         if (op == TOK_UMULL) {
 785             o(0xf7); /* mul fr */
 786             o(0xe0 + fr);
 787             vtop->r2 = TREG_EDX;
 788             r = TREG_EAX;
 789         } else {
 790             if (op == TOK_UDIV || op == TOK_UMOD) {
 791                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 792                 o(0xf0 + fr);
 793             } else {
 794                 o(0xf799); /* cltd, idiv fr, %eax */
 795                 o(0xf8 + fr);
 796             }
 797             if (op == '%' || op == TOK_UMOD)
 798                 r = TREG_EDX;
 799             else
 800                 r = TREG_EAX;
 801         }
 802         vtop->r = r;
 803         break;
 804     default:
 805         opc = 7;
 806         goto gen_op8;
 807     }
 808 }
 809
 810 /* generate a floating point operation 'v = t1 op t2' instruction. The
 811    two operands are guaranted to have the same floating point type */
 812 /* XXX: need to use ST1 too */
 813 ST_FUNC void gen_opf(int op)
 814 {
 815     int a, ft, fc, swapped, r;
 816
 817     /* convert constants to memory references */
 818     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 819         vswap();
 820         gv(RC_FLOAT);
 821         vswap();
 822     }
 823     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 824         gv(RC_FLOAT);
 825
 826     /* must put at least one value in the floating point register */
 827     if ((vtop[-1].r & VT_LVAL) &&
 828         (vtop[0].r & VT_LVAL)) {
 829         vswap();
 830         gv(RC_FLOAT);
 831         vswap();
 832     }
 833     swapped = 0;
 834     /* swap the stack if needed so that t1 is the register and t2 is
 835        the memory reference */
 836     if (vtop[-1].r & VT_LVAL) {
 837         vswap();
 838         swapped = 1;
 839     }
 840     if (op >= TOK_ULT && op <= TOK_GT) {
 841         /* load on stack second operand */
 842         load(TREG_ST0, vtop);
 843         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 844         if (op == TOK_GE || op == TOK_GT)
 845             swapped = !swapped;
 846         else if (op == TOK_EQ || op == TOK_NE)
 847             swapped = 0;
 848         if (swapped)
 849             o(0xc9d9); /* fxch %st(1) */
 850         o(0xe9da); /* fucompp */
 851         o(0xe0df); /* fnstsw %ax */
 852         if (op == TOK_EQ) {
 853             o(0x45e480); /* and $0x45, %ah */
 854             o(0x40fC80); /* cmp $0x40, %ah */
 855         } else if (op == TOK_NE) {
 856             o(0x45e480); /* and $0x45, %ah */
 857             o(0x40f480); /* xor $0x40, %ah */
 858             op = TOK_NE;
 859         } else if (op == TOK_GE || op == TOK_LE) {
 860             o(0x05c4f6); /* test $0x05, %ah */
 861             op = TOK_EQ;
 862         } else {
 863             o(0x45c4f6); /* test $0x45, %ah */
 864             op = TOK_EQ;
 865         }
 866         vtop--;
 867         vtop->r = VT_CMP;
 868         vtop->c.i = op;
 869     } else {
 870         /* no memory reference possible for long double operations */
 871         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 872             load(TREG_ST0, vtop);
 873             swapped = !swapped;
 874         }
 875
 876         switch(op) {
 877         default:
 878         case '+':
 879             a = 0;
 880             break;
 881         case '-':
 882             a = 4;
 883             if (swapped)
 884                 a++;
 885             break;
 886         case '*':
 887             a = 1;
 888             break;
 889         case '/':
 890             a = 6;
 891             if (swapped)
 892                 a++;
 893             break;
 894         }
 895         ft = vtop->type.t;
 896         fc = vtop->c.ul;
 897         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 898             o(0xde); /* fxxxp %st, %st(1) */
 899             o(0xc1 + (a << 3));
 900         } else {
 901             /* if saved lvalue, then we must reload it */
 902             r = vtop->r;
 903             if ((r & VT_VALMASK) == VT_LLOCAL) {
 904                 SValue v1;
 905                 r = get_reg(RC_INT);
 906                 v1.type.t = VT_INT;
 907                 v1.r = VT_LOCAL | VT_LVAL;
 908                 v1.c.ul = fc;
 909                 load(r, &v1);
 910                 fc = 0;
 911             }
 912
 913             if ((ft & VT_BTYPE) == VT_DOUBLE)
 914                 o(0xdc);
 915             else
 916                 o(0xd8);
 917             gen_modrm(a, r, vtop->sym, fc);
 918         }
 919         vtop--;
 920     }
 921 }
 922
 923 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 924    and 'long long' cases. */
 925 ST_FUNC void gen_cvt_itof(int t)
 926 {
 927     save_reg(TREG_ST0);
 928     gv(RC_INT);
 929     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 930         /* signed long long to float/double/long double (unsigned case
 931            is handled generically) */
 932         o(0x50 + vtop->r2); /* push r2 */
 933         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 934         o(0x242cdf); /* fildll (%esp) */
 935         o(0x08c483); /* add $8, %esp */
 936     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 937                (VT_INT | VT_UNSIGNED)) {
 938         /* unsigned int to float/double/long double */
 939         o(0x6a); /* push $0 */
 940         g(0x00);
 941         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 942         o(0x242cdf); /* fildll (%esp) */
 943         o(0x08c483); /* add $8, %esp */
 944     } else {
 945         /* int to float/double/long double */
 946         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 947         o(0x2404db); /* fildl (%esp) */
 948         o(0x04c483); /* add $4, %esp */
 949     }
 950     vtop->r = TREG_ST0;
 951 }
 952
 953 /* convert fp to int 't' type */
 954 /* XXX: handle long long case */
 955 ST_FUNC void gen_cvt_ftoi(int t)
 956 {
 957     int r, r2, size;
 958     Sym *sym;
 959     CType ushort_type;
 960
 961     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 962     ushort_type.ref = 0;
 963
 964     gv(RC_FLOAT);
 965     if (t != VT_INT)
 966         size = 8;
 967     else
 968         size = 4;
 969
 970     o(0x2dd9); /* ldcw xxx */
 971     sym = external_global_sym(TOK___tcc_int_fpu_control,
 972                               &ushort_type, VT_LVAL);
 973     greloc(cur_text_section, sym,
 974            ind, R_386_32);
 975     gen_le32(0);
 976
 977     oad(0xec81, size); /* sub $xxx, %esp */
 978     if (size == 4)
 979         o(0x1cdb); /* fistpl */
 980     else
 981         o(0x3cdf); /* fistpll */
 982     o(0x24);
 983     o(0x2dd9); /* ldcw xxx */
 984     sym = external_global_sym(TOK___tcc_fpu_control,
 985                               &ushort_type, VT_LVAL);
 986     greloc(cur_text_section, sym,
 987            ind, R_386_32);
 988     gen_le32(0);
 989
 990     r = get_reg(RC_INT);
 991     o(0x58 + r); /* pop r */
 992     if (size == 8) {
 993         if (t == VT_LLONG) {
 994             vtop->r = r; /* mark reg as used */
 995             r2 = get_reg(RC_INT);
 996             o(0x58 + r2); /* pop r2 */
 997             vtop->r2 = r2;
 998         } else {
 999             o(0x04c483); /* add $4, %esp */
1000         }
1001     }
1002     vtop->r = r;
1003 }
1004
1005 /* convert from one floating point type to another */
1006 ST_FUNC void gen_cvt_ftof(int t)
1007 {
1008     /* all we have to do on i386 is to put the float in a register */
1009     gv(RC_FLOAT);
1010 }
1011
1012 /* computed goto support */
1013 ST_FUNC void ggoto(void)
1014 {
1015     gcall_or_jmp(1);
1016     vtop--;
1017 }
1018
1019 /* bound check support functions */
1020 #ifdef CONFIG_TCC_BCHECK
1021
1022 /* generate a bounded pointer addition */
1023 ST_FUNC void gen_bounded_ptr_add(void)
1024 {
1025     Sym *sym;
1026
1027     /* prepare fast i386 function call (args in eax and edx) */
1028     gv2(RC_EAX, RC_EDX);
1029     /* save all temporary registers */
1030     vtop -= 2;
1031     save_regs(0);
1032     /* do a fast function call */
1033     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
1034     greloc(cur_text_section, sym,
1035            ind + 1, R_386_PC32);
1036     oad(0xe8, -4);
1037     /* returned pointer is in eax */
1038     vtop++;
1039     vtop->r = TREG_EAX | VT_BOUNDED;
1040     /* address of bounding function call point */
1041     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1042 }
1043
1044 /* patch pointer addition in vtop so that pointer dereferencing is
1045    also tested */
1046 ST_FUNC void gen_bounded_ptr_deref(void)
1047 {
1048     int func;
1049     int size, align;
1050     Elf32_Rel *rel;
1051     Sym *sym;
1052
1053     size = 0;
1054     /* XXX: put that code in generic part of tcc */
1055     if (!is_float(vtop->type.t)) {
1056         if (vtop->r & VT_LVAL_BYTE)
1057             size = 1;
1058         else if (vtop->r & VT_LVAL_SHORT)
1059             size = 2;
1060     }
1061     if (!size)
1062         size = type_size(&vtop->type, &align);
1063     switch(size) {
1064     case  1: func = TOK___bound_ptr_indir1; break;
1065     case  2: func = TOK___bound_ptr_indir2; break;
1066     case  4: func = TOK___bound_ptr_indir4; break;
1067     case  8: func = TOK___bound_ptr_indir8; break;
1068     case 12: func = TOK___bound_ptr_indir12; break;
1069     case 16: func = TOK___bound_ptr_indir16; break;
1070     default:
1071         error("unhandled size when derefencing bounded pointer");
1072         func = 0;
1073         break;
1074     }
1075
1076     /* patch relocation */
1077     /* XXX: find a better solution ? */
1078     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1079     sym = external_global_sym(func, &func_old_type, 0);
1080     if (!sym->c)
1081         put_extern_sym(sym, NULL, 0, 0);
1082     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1083 }
1084 #endif
1085
1086 /* end of X86 code generator */
1087 /*************************************************************/
1088 #endif
1089 /*************************************************************/