i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001, 2002, 2003 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 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 /* relocation type for 32 bit data relocation */
  74 #define R_DATA_32 R_386_32
  75
  76 /******************************************************/
  77
  78 static unsigned long func_sub_sp_offset;
  79 static unsigned long func_bound_offset;
  80 static int func_ret_sub;
  81
  82 /* XXX: make it faster ? */
  83 void g(int c)
  84 {
  85     int ind1;
  86     ind1 = ind + 1;
  87     if (ind1 > cur_text_section->data_allocated)
  88         section_realloc(cur_text_section, ind1);
  89     cur_text_section->data[ind] = c;
  90     ind = ind1;
  91 }
  92
  93 void o(int c)
  94 {
  95     while (c) {
  96         g(c);
  97         c = c / 256;
  98     }
  99 }
 100
 101 void gen_le32(int c)
 102 {
 103     g(c);
 104     g(c >> 8);
 105     g(c >> 16);
 106     g(c >> 24);
 107 }
 108
 109 /* output a symbol and patch all calls to it */
 110 void gsym_addr(int t, int a)
 111 {
 112     int n, *ptr;
 113     while (t) {
 114         ptr = (int *)(cur_text_section->data + t);
 115         n = *ptr; /* next value */
 116         *ptr = a - t - 4;
 117         t = n;
 118     }
 119 }
 120
 121 void gsym(int t)
 122 {
 123     gsym_addr(t, ind);
 124 }
 125
 126 /* psym is used to put an instruction with a data field which is a
 127    reference to a symbol. It is in fact the same as oad ! */
 128 #define psym oad
 129
 130 /* instruction + 4 bytes data. Return the address of the data */
 131 static int oad(int c, int s)
 132 {
 133     int ind1;
 134
 135     o(c);
 136     ind1 = ind + 4;
 137     if (ind1 > cur_text_section->data_allocated)
 138         section_realloc(cur_text_section, ind1);
 139     *(int *)(cur_text_section->data + ind) = s;
 140     s = ind;
 141     ind = ind1;
 142     return s;
 143 }
 144
 145 /* output constant with relocation if 'r & VT_SYM' is true */
 146 static void gen_addr32(int r, Sym *sym, int c)
 147 {
 148     if (r & VT_SYM)
 149         greloc(cur_text_section, sym, ind, R_386_32);
 150     gen_le32(c);
 151 }
 152
 153 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 154    opcode bits */
 155 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 156 {
 157     op_reg = op_reg << 3;
 158     if ((r & VT_VALMASK) == VT_CONST) {
 159         /* constant memory reference */
 160         o(0x05 | op_reg);
 161         gen_addr32(r, sym, c);
 162     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 163         /* currently, we use only ebp as base */
 164         if (c == (char)c) {
 165             /* short reference */
 166             o(0x45 | op_reg);
 167             g(c);
 168         } else {
 169             oad(0x85 | op_reg, c);
 170         }
 171     } else {
 172         g(0x00 | op_reg | (r & VT_VALMASK));
 173     }
 174 }
 175
 176
 177 /* load 'r' from value 'sv' */
 178 void load(int r, SValue *sv)
 179 {
 180     int v, t, ft, fc, fr;
 181     SValue v1;
 182
 183     fr = sv->r;
 184     ft = sv->type.t;
 185     fc = sv->c.ul;
 186
 187     v = fr & VT_VALMASK;
 188     if (fr & VT_LVAL) {
 189         if (v == VT_LLOCAL) {
 190             v1.type.t = VT_INT;
 191             v1.r = VT_LOCAL | VT_LVAL;
 192             v1.c.ul = fc;
 193             load(r, &v1);
 194             fr = r;
 195         }
 196         if ((ft & VT_BTYPE) == VT_FLOAT) {
 197             o(0xd9); /* flds */
 198             r = 0;
 199         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 200             o(0xdd); /* fldl */
 201             r = 0;
 202         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 203             o(0xdb); /* fldt */
 204             r = 5;
 205         } else if ((ft & VT_TYPE) == VT_BYTE) {
 206             o(0xbe0f);   /* movsbl */
 207         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 208             o(0xb60f);   /* movzbl */
 209         } else if ((ft & VT_TYPE) == VT_SHORT) {
 210             o(0xbf0f);   /* movswl */
 211         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 212             o(0xb70f);   /* movzwl */
 213         } else {
 214             o(0x8b);     /* movl */
 215         }
 216         gen_modrm(r, fr, sv->sym, fc);
 217     } else {
 218         if (v == VT_CONST) {
 219             o(0xb8 + r); /* mov $xx, r */
 220             gen_addr32(fr, sv->sym, fc);
 221         } else if (v == VT_LOCAL) {
 222             o(0x8d); /* lea xxx(%ebp), r */
 223             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 224         } else if (v == VT_CMP) {
 225             oad(0xb8 + r, 0); /* mov $0, r */
 226             o(0x0f); /* setxx %br */
 227             o(fc);
 228             o(0xc0 + r);
 229         } else if (v == VT_JMP || v == VT_JMPI) {
 230             t = v & 1;
 231             oad(0xb8 + r, t); /* mov $1, r */
 232             o(0x05eb); /* jmp after */
 233             gsym(fc);
 234             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 235         } else if (v != r) {
 236             o(0x89);
 237             o(0xc0 + r + v * 8); /* mov v, r */
 238         }
 239     }
 240 }
 241
 242 /* store register 'r' in lvalue 'v' */
 243 void store(int r, SValue *v)
 244 {
 245     int fr, bt, ft, fc;
 246
 247     ft = v->type.t;
 248     fc = v->c.ul;
 249     fr = v->r & VT_VALMASK;
 250     bt = ft & VT_BTYPE;
 251     /* XXX: incorrect if float reg to reg */
 252     if (bt == VT_FLOAT) {
 253         o(0xd9); /* fsts */
 254         r = 2;
 255     } else if (bt == VT_DOUBLE) {
 256         o(0xdd); /* fstpl */
 257         r = 2;
 258     } else if (bt == VT_LDOUBLE) {
 259         o(0xc0d9); /* fld %st(0) */
 260         o(0xdb); /* fstpt */
 261         r = 7;
 262     } else {
 263         if (bt == VT_SHORT)
 264             o(0x66);
 265         if (bt == VT_BYTE)
 266             o(0x88);
 267         else
 268             o(0x89);
 269     }
 270     if (fr == VT_CONST ||
 271         fr == VT_LOCAL ||
 272         (v->r & VT_LVAL)) {
 273         gen_modrm(r, v->r, v->sym, fc);
 274     } else if (fr != r) {
 275         o(0xc0 + fr + r * 8); /* mov r, fr */
 276     }
 277 }
 278
 279 static void gadd_sp(int val)
 280 {
 281     if (val == (char)val) {
 282         o(0xc483);
 283         g(val);
 284     } else {
 285         oad(0xc481, val); /* add $xxx, %esp */
 286     }
 287 }
 288
 289 /* 'is_jmp' is '1' if it is a jump */
 290 static void gcall_or_jmp(int is_jmp)
 291 {
 292     int r;
 293     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 294         /* constant case */
 295         if (vtop->r & VT_SYM) {
 296             /* relocation case */
 297             greloc(cur_text_section, vtop->sym,
 298                    ind + 1, R_386_PC32);
 299         } else {
 300             /* put an empty PC32 relocation */
 301             put_elf_reloc(symtab_section, cur_text_section,
 302                           ind + 1, R_386_PC32, 0);
 303         }
 304         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 305     } else {
 306         /* otherwise, indirect call */
 307         r = gv(RC_INT);
 308         o(0xff); /* call/jmp *r */
 309         o(0xd0 + r + (is_jmp << 4));
 310     }
 311 }
 312
 313 /* Generate function call. The function address is pushed first, then
 314    all the parameters in call order. This functions pops all the
 315    parameters and the function address. */
 316 void gfunc_call(int nb_args)
 317 {
 318     int size, align, r, args_size, i;
 319     Sym *func_sym;
 320
 321     args_size = 0;
 322     for(i = 0;i < nb_args; i++) {
 323         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 324             size = type_size(&vtop->type, &align);
 325             /* align to stack align size */
 326             size = (size + 3) & ~3;
 327             /* allocate the necessary size on stack */
 328             oad(0xec81, size); /* sub $xxx, %esp */
 329             /* generate structure store */
 330             r = get_reg(RC_INT);
 331             o(0x89); /* mov %esp, r */
 332             o(0xe0 + r);
 333             vset(&vtop->type, r | VT_LVAL, 0);
 334             vswap();
 335             vstore();
 336             args_size += size;
 337         } else if (is_float(vtop->type.t)) {
 338             gv(RC_FLOAT); /* only one float register */
 339             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 340                 size = 4;
 341             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 342                 size = 8;
 343             else
 344                 size = 12;
 345             oad(0xec81, size); /* sub $xxx, %esp */
 346             if (size == 12)
 347                 o(0x7cdb);
 348             else
 349                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 350             g(0x24);
 351             g(0x00);
 352             args_size += size;
 353         } else {
 354             /* simple type (currently always same size) */
 355             /* XXX: implicit cast ? */
 356             r = gv(RC_INT);
 357             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 358                 size = 8;
 359                 o(0x50 + vtop->r2); /* push r */
 360             } else {
 361                 size = 4;
 362             }
 363             o(0x50 + r); /* push r */
 364             args_size += size;
 365         }
 366         vtop--;
 367     }
 368     save_regs(0); /* save used temporary registers */
 369     func_sym = vtop->type.ref;
 370     gcall_or_jmp(0);
 371     if (args_size && func_sym->r == FUNC_CDECL)
 372         gadd_sp(args_size);
 373     vtop--;
 374 }
 375
 376 /* generate function prolog of type 't' */
 377 void gfunc_prolog(CType *func_type)
 378 {
 379     int addr, align, size, func_call;
 380     Sym *sym;
 381     CType *type;
 382
 383     sym = func_type->ref;
 384     func_call = sym->r;
 385     addr = 8;
 386     /* if the function returns a structure, then add an
 387        implicit pointer parameter */
 388     func_vt = sym->type;
 389     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 390         func_vc = addr;
 391         addr += 4;
 392     }
 393     /* define parameters */
 394     while ((sym = sym->next) != NULL) {
 395         type = &sym->type;
 396         sym_push(sym->v & ~SYM_FIELD, type,
 397                  VT_LOCAL | VT_LVAL, addr);
 398         size = type_size(type, &align);
 399         size = (size + 3) & ~3;
 400 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 401         /* structs are passed as pointer */
 402         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 403             size = 4;
 404         }
 405 #endif
 406         addr += size;
 407     }
 408     func_ret_sub = 0;
 409     /* pascal type call ? */
 410     if (func_call == FUNC_STDCALL)
 411         func_ret_sub = addr - 8;
 412     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 413     func_sub_sp_offset = oad(0xec81, 0); /* sub $xxx, %esp */
 414     /* leave some room for bound checking code */
 415     if (do_bounds_check) {
 416         oad(0xb8, 0); /* lbound section pointer */
 417         oad(0xb8, 0); /* call to function */
 418         func_bound_offset = lbounds_section->data_offset;
 419     }
 420 }
 421
 422 /* generate function epilog */
 423 void gfunc_epilog(void)
 424 {
 425 #ifdef CONFIG_TCC_BCHECK
 426     if (do_bounds_check && func_bound_offset != lbounds_section->data_offset) {
 427         int saved_ind;
 428         int *bounds_ptr;
 429         Sym *sym, *sym_data;
 430         /* add end of table info */
 431         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 432         *bounds_ptr = 0;
 433         /* generate bound local allocation */
 434         saved_ind = ind;
 435         ind = func_sub_sp_offset + 4;
 436         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 437                                func_bound_offset, lbounds_section->data_offset);
 438         greloc(cur_text_section, sym_data,
 439                ind + 1, R_386_32);
 440         oad(0xb8, 0); /* mov %eax, xxx */
 441         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 442         greloc(cur_text_section, sym,
 443                ind + 1, R_386_PC32);
 444         oad(0xe8, -4);
 445         ind = saved_ind;
 446         /* generate bound check local freeing */
 447         o(0x5250); /* save returned value, if any */
 448         greloc(cur_text_section, sym_data,
 449                ind + 1, R_386_32);
 450         oad(0xb8, 0); /* mov %eax, xxx */
 451         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 452         greloc(cur_text_section, sym,
 453                ind + 1, R_386_PC32);
 454         oad(0xe8, -4);
 455         o(0x585a); /* restore returned value, if any */
 456     }
 457 #endif
 458     o(0xc9); /* leave */
 459     if (func_ret_sub == 0) {
 460         o(0xc3); /* ret */
 461     } else {
 462         o(0xc2); /* ret n */
 463         g(func_ret_sub);
 464         g(func_ret_sub >> 8);
 465     }
 466     /* align local size to word & save local variables */
 467     *(int *)(cur_text_section->data + func_sub_sp_offset) = (-loc + 3) & -4;
 468 }
 469
 470 /* generate a jump to a label */
 471 int gjmp(int t)
 472 {
 473     return psym(0xe9, t);
 474 }
 475
 476 /* generate a jump to a fixed address */
 477 void gjmp_addr(int a)
 478 {
 479     int r;
 480     r = a - ind - 2;
 481     if (r == (char)r) {
 482         g(0xeb);
 483         g(r);
 484     } else {
 485         oad(0xe9, a - ind - 5);
 486     }
 487 }
 488
 489 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 490 int gtst(int inv, int t)
 491 {
 492     int v, *p;
 493
 494     v = vtop->r & VT_VALMASK;
 495     if (v == VT_CMP) {
 496         /* fast case : can jump directly since flags are set */
 497         g(0x0f);
 498         t = psym((vtop->c.i - 16) ^ inv, t);
 499     } else if (v == VT_JMP || v == VT_JMPI) {
 500         /* && or || optimization */
 501         if ((v & 1) == inv) {
 502             /* insert vtop->c jump list in t */
 503             p = &vtop->c.i;
 504             while (*p != 0)
 505                 p = (int *)(cur_text_section->data + *p);
 506             *p = t;
 507             t = vtop->c.i;
 508         } else {
 509             t = gjmp(t);
 510             gsym(vtop->c.i);
 511         }
 512     } else {
 513         if (is_float(vtop->type.t)) {
 514             vpushi(0);
 515             gen_op(TOK_NE);
 516         }
 517         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 518             /* constant jmp optimization */
 519             if ((vtop->c.i != 0) != inv)
 520                 t = gjmp(t);
 521         } else {
 522             v = gv(RC_INT);
 523             o(0x85);
 524             o(0xc0 + v * 9);
 525             g(0x0f);
 526             t = psym(0x85 ^ inv, t);
 527         }
 528     }
 529     vtop--;
 530     return t;
 531 }
 532
 533 /* generate an integer binary operation */
 534 void gen_opi(int op)
 535 {
 536     int r, fr, opc, c;
 537
 538     switch(op) {
 539     case '+':
 540     case TOK_ADDC1: /* add with carry generation */
 541         opc = 0;
 542     gen_op8:
 543         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 544             /* constant case */
 545             vswap();
 546             r = gv(RC_INT);
 547             vswap();
 548             c = vtop->c.i;
 549             if (c == (char)c) {
 550                 /* XXX: generate inc and dec for smaller code ? */
 551                 o(0x83);
 552                 o(0xc0 | (opc << 3) | r);
 553                 g(c);
 554             } else {
 555                 o(0x81);
 556                 oad(0xc0 | (opc << 3) | r, c);
 557             }
 558         } else {
 559             gv2(RC_INT, RC_INT);
 560             r = vtop[-1].r;
 561             fr = vtop[0].r;
 562             o((opc << 3) | 0x01);
 563             o(0xc0 + r + fr * 8);
 564         }
 565         vtop--;
 566         if (op >= TOK_ULT && op <= TOK_GT) {
 567             vtop->r = VT_CMP;
 568             vtop->c.i = op;
 569         }
 570         break;
 571     case '-':
 572     case TOK_SUBC1: /* sub with carry generation */
 573         opc = 5;
 574         goto gen_op8;
 575     case TOK_ADDC2: /* add with carry use */
 576         opc = 2;
 577         goto gen_op8;
 578     case TOK_SUBC2: /* sub with carry use */
 579         opc = 3;
 580         goto gen_op8;
 581     case '&':
 582         opc = 4;
 583         goto gen_op8;
 584     case '^':
 585         opc = 6;
 586         goto gen_op8;
 587     case '|':
 588         opc = 1;
 589         goto gen_op8;
 590     case '*':
 591         gv2(RC_INT, RC_INT);
 592         r = vtop[-1].r;
 593         fr = vtop[0].r;
 594         vtop--;
 595         o(0xaf0f); /* imul fr, r */
 596         o(0xc0 + fr + r * 8);
 597         break;
 598     case TOK_SHL:
 599         opc = 4;
 600         goto gen_shift;
 601     case TOK_SHR:
 602         opc = 5;
 603         goto gen_shift;
 604     case TOK_SAR:
 605         opc = 7;
 606     gen_shift:
 607         opc = 0xc0 | (opc << 3);
 608         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 609             /* constant case */
 610             vswap();
 611             r = gv(RC_INT);
 612             vswap();
 613             c = vtop->c.i & 0x1f;
 614             o(0xc1); /* shl/shr/sar $xxx, r */
 615             o(opc | r);
 616             g(c);
 617         } else {
 618             /* we generate the shift in ecx */
 619             gv2(RC_INT, RC_ECX);
 620             r = vtop[-1].r;
 621             o(0xd3); /* shl/shr/sar %cl, r */
 622             o(opc | r);
 623         }
 624         vtop--;
 625         break;
 626     case '/':
 627     case TOK_UDIV:
 628     case TOK_PDIV:
 629     case '%':
 630     case TOK_UMOD:
 631     case TOK_UMULL:
 632         /* first operand must be in eax */
 633         /* XXX: need better constraint for second operand */
 634         gv2(RC_EAX, RC_ECX);
 635         r = vtop[-1].r;
 636         fr = vtop[0].r;
 637         vtop--;
 638         save_reg(TREG_EDX);
 639         if (op == TOK_UMULL) {
 640             o(0xf7); /* mul fr */
 641             o(0xe0 + fr);
 642             vtop->r2 = TREG_EDX;
 643             r = TREG_EAX;
 644         } else {
 645             if (op == TOK_UDIV || op == TOK_UMOD) {
 646                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 647                 o(0xf0 + fr);
 648             } else {
 649                 o(0xf799); /* cltd, idiv fr, %eax */
 650                 o(0xf8 + fr);
 651             }
 652             if (op == '%' || op == TOK_UMOD)
 653                 r = TREG_EDX;
 654             else
 655                 r = TREG_EAX;
 656         }
 657         vtop->r = r;
 658         break;
 659     default:
 660         opc = 7;
 661         goto gen_op8;
 662     }
 663 }
 664
 665 /* generate a floating point operation 'v = t1 op t2' instruction. The
 666    two operands are guaranted to have the same floating point type */
 667 /* XXX: need to use ST1 too */
 668 void gen_opf(int op)
 669 {
 670     int a, ft, fc, swapped, r;
 671
 672     /* convert constants to memory references */
 673     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 674         vswap();
 675         gv(RC_FLOAT);
 676         vswap();
 677     }
 678     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 679         gv(RC_FLOAT);
 680
 681     /* must put at least one value in the floating point register */
 682     if ((vtop[-1].r & VT_LVAL) &&
 683         (vtop[0].r & VT_LVAL)) {
 684         vswap();
 685         gv(RC_FLOAT);
 686         vswap();
 687     }
 688     swapped = 0;
 689     /* swap the stack if needed so that t1 is the register and t2 is
 690        the memory reference */
 691     if (vtop[-1].r & VT_LVAL) {
 692         vswap();
 693         swapped = 1;
 694     }
 695     if (op >= TOK_ULT && op <= TOK_GT) {
 696         /* load on stack second operand */
 697         load(TREG_ST0, vtop);
 698         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 699         if (op == TOK_GE || op == TOK_GT)
 700             swapped = !swapped;
 701         else if (op == TOK_EQ || op == TOK_NE)
 702             swapped = 0;
 703         if (swapped)
 704             o(0xc9d9); /* fxch %st(1) */
 705         o(0xe9da); /* fucompp */
 706         o(0xe0df); /* fnstsw %ax */
 707         if (op == TOK_EQ) {
 708             o(0x45e480); /* and $0x45, %ah */
 709             o(0x40fC80); /* cmp $0x40, %ah */
 710         } else if (op == TOK_NE) {
 711             o(0x45e480); /* and $0x45, %ah */
 712             o(0x40f480); /* xor $0x40, %ah */
 713             op = TOK_NE;
 714         } else if (op == TOK_GE || op == TOK_LE) {
 715             o(0x05c4f6); /* test $0x05, %ah */
 716             op = TOK_EQ;
 717         } else {
 718             o(0x45c4f6); /* test $0x45, %ah */
 719             op = TOK_EQ;
 720         }
 721         vtop--;
 722         vtop->r = VT_CMP;
 723         vtop->c.i = op;
 724     } else {
 725         /* no memory reference possible for long double operations */
 726         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 727             load(TREG_ST0, vtop);
 728             swapped = !swapped;
 729         }
 730
 731         switch(op) {
 732         default:
 733         case '+':
 734             a = 0;
 735             break;
 736         case '-':
 737             a = 4;
 738             if (swapped)
 739                 a++;
 740             break;
 741         case '*':
 742             a = 1;
 743             break;
 744         case '/':
 745             a = 6;
 746             if (swapped)
 747                 a++;
 748             break;
 749         }
 750         ft = vtop->type.t;
 751         fc = vtop->c.ul;
 752         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 753             o(0xde); /* fxxxp %st, %st(1) */
 754             o(0xc1 + (a << 3));
 755         } else {
 756             /* if saved lvalue, then we must reload it */
 757             r = vtop->r;
 758             if ((r & VT_VALMASK) == VT_LLOCAL) {
 759                 SValue v1;
 760                 r = get_reg(RC_INT);
 761                 v1.type.t = VT_INT;
 762                 v1.r = VT_LOCAL | VT_LVAL;
 763                 v1.c.ul = fc;
 764                 load(r, &v1);
 765                 fc = 0;
 766             }
 767
 768             if ((ft & VT_BTYPE) == VT_DOUBLE)
 769                 o(0xdc);
 770             else
 771                 o(0xd8);
 772             gen_modrm(a, r, vtop->sym, fc);
 773         }
 774         vtop--;
 775     }
 776 }
 777
 778 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 779    and 'long long' cases. */
 780 void gen_cvt_itof(int t)
 781 {
 782     save_reg(TREG_ST0);
 783     gv(RC_INT);
 784     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 785         /* signed long long to float/double/long double (unsigned case
 786            is handled generically) */
 787         o(0x50 + vtop->r2); /* push r2 */
 788         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 789         o(0x242cdf); /* fildll (%esp) */
 790         o(0x08c483); /* add $8, %esp */
 791     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 792                (VT_INT | VT_UNSIGNED)) {
 793         /* unsigned int to float/double/long double */
 794         o(0x6a); /* push $0 */
 795         g(0x00);
 796         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 797         o(0x242cdf); /* fildll (%esp) */
 798         o(0x08c483); /* add $8, %esp */
 799     } else {
 800         /* int to float/double/long double */
 801         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 802         o(0x2404db); /* fildl (%esp) */
 803         o(0x04c483); /* add $4, %esp */
 804     }
 805     vtop->r = TREG_ST0;
 806 }
 807
 808 /* convert fp to int 't' type */
 809 /* XXX: handle long long case */
 810 void gen_cvt_ftoi(int t)
 811 {
 812     int r, r2, size;
 813     Sym *sym;
 814     CType ushort_type;
 815
 816     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 817
 818     gv(RC_FLOAT);
 819     if (t != VT_INT)
 820         size = 8;
 821     else
 822         size = 4;
 823
 824     o(0x2dd9); /* ldcw xxx */
 825     sym = external_global_sym(TOK___tcc_int_fpu_control,
 826                               &ushort_type, VT_LVAL);
 827     greloc(cur_text_section, sym,
 828            ind, R_386_32);
 829     gen_le32(0);
 830
 831     oad(0xec81, size); /* sub $xxx, %esp */
 832     if (size == 4)
 833         o(0x1cdb); /* fistpl */
 834     else
 835         o(0x3cdf); /* fistpll */
 836     o(0x24);
 837     o(0x2dd9); /* ldcw xxx */
 838     sym = external_global_sym(TOK___tcc_fpu_control,
 839                               &ushort_type, VT_LVAL);
 840     greloc(cur_text_section, sym,
 841            ind, R_386_32);
 842     gen_le32(0);
 843
 844     r = get_reg(RC_INT);
 845     o(0x58 + r); /* pop r */
 846     if (size == 8) {
 847         if (t == VT_LLONG) {
 848             vtop->r = r; /* mark reg as used */
 849             r2 = get_reg(RC_INT);
 850             o(0x58 + r2); /* pop r2 */
 851             vtop->r2 = r2;
 852         } else {
 853             o(0x04c483); /* add $4, %esp */
 854         }
 855     }
 856     vtop->r = r;
 857 }
 858
 859 /* convert from one floating point type to another */
 860 void gen_cvt_ftof(int t)
 861 {
 862     /* all we have to do on i386 is to put the float in a register */
 863     gv(RC_FLOAT);
 864 }
 865
 866 /* computed goto support */
 867 void ggoto(void)
 868 {
 869     gcall_or_jmp(1);
 870     vtop--;
 871 }
 872
 873 /* bound check support functions */
 874 #ifdef CONFIG_TCC_BCHECK
 875
 876 /* generate a bounded pointer addition */
 877 void gen_bounded_ptr_add(void)
 878 {
 879     Sym *sym;
 880
 881     /* prepare fast i386 function call (args in eax and edx) */
 882     gv2(RC_EAX, RC_EDX);
 883     /* save all temporary registers */
 884     vtop -= 2;
 885     save_regs(0);
 886     /* do a fast function call */
 887     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
 888     greloc(cur_text_section, sym,
 889            ind + 1, R_386_PC32);
 890     oad(0xe8, -4);
 891     /* returned pointer is in eax */
 892     vtop++;
 893     vtop->r = TREG_EAX | VT_BOUNDED;
 894     /* address of bounding function call point */
 895     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 896 }
 897
 898 /* patch pointer addition in vtop so that pointer dereferencing is
 899    also tested */
 900 void gen_bounded_ptr_deref(void)
 901 {
 902     int func;
 903     int size, align;
 904     Elf32_Rel *rel;
 905     Sym *sym;
 906
 907     size = 0;
 908     /* XXX: put that code in generic part of tcc */
 909     if (!is_float(vtop->type.t)) {
 910         if (vtop->r & VT_LVAL_BYTE)
 911             size = 1;
 912         else if (vtop->r & VT_LVAL_SHORT)
 913             size = 2;
 914     }
 915     if (!size)
 916         size = type_size(&vtop->type, &align);
 917     switch(size) {
 918     case  1: func = TOK___bound_ptr_indir1; break;
 919     case  2: func = TOK___bound_ptr_indir2; break;
 920     case  4: func = TOK___bound_ptr_indir4; break;
 921     case  8: func = TOK___bound_ptr_indir8; break;
 922     case 12: func = TOK___bound_ptr_indir12; break;
 923     case 16: func = TOK___bound_ptr_indir16; break;
 924     default:
 925         error("unhandled size when derefencing bounded pointer");
 926         func = 0;
 927         break;
 928     }
 929
 930     /* patch relocation */
 931     /* XXX: find a better solution ? */
 932     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
 933     sym = external_global_sym(func, &func_old_type, 0);
 934     if (!sym->c)
 935         put_extern_sym(sym, NULL, 0, 0);
 936     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
 937 }
 938 #endif
 939
 940 /* end of X86 code generator */
 941 /*************************************************************/
 942