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