i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001, 2002 Fabrice Bellard
   5  *
   6  *  This program is free software; you can redistribute it and/or modify
   7  *  it under the terms of the GNU General Public License as published by
   8  *  the Free Software Foundation; either version 2 of the License, or
   9  *  (at your option) any later version.
  10  *
  11  *  This program 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
  14  *  GNU General Public License for more details.
  15  *
  16  *  You should have received a copy of the GNU General Public License
  17  *  along with this program; if not, write to the Free Software
  18  *  Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, 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     REG_EAX = 0,
  40     REG_ECX,
  41     REG_EDX,
  42     REG_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 REG_EAX /* single word int return register */
  54 #define REG_LRET REG_EDX /* second word return register (for long long) */
  55 #define REG_FRET REG_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
  71 /* relocation type for 32 bit data relocation */
  72 #define R_DATA_32 R_386_32
  73
  74 /* function call context */
  75 typedef struct GFuncContext {
  76     int args_size;
  77     int func_call; /* func call type (FUNC_STDCALL or FUNC_CDECL) */
  78 } GFuncContext;
  79
  80 /******************************************************/
  81
  82 static unsigned long func_sub_sp_offset;
  83 static unsigned long func_bound_offset;
  84 static int func_ret_sub;
  85
  86 /* XXX: make it faster ? */
  87 void g(int c)
  88 {
  89     int ind1;
  90     ind1 = ind + 1;
  91     if (ind1 > cur_text_section->data_allocated)
  92         section_realloc(cur_text_section, ind1);
  93     cur_text_section->data[ind] = c;
  94     ind = ind1;
  95 }
  96
  97 void o(int c)
  98 {
  99     while (c) {
 100         g(c);
 101         c = c / 256;
 102     }
 103 }
 104
 105 void gen_le32(int c)
 106 {
 107     g(c);
 108     g(c >> 8);
 109     g(c >> 16);
 110     g(c >> 24);
 111 }
 112
 113 /* output a symbol and patch all calls to it */
 114 void gsym_addr(int t, int a)
 115 {
 116     int n, *ptr;
 117     while (t) {
 118         ptr = (int *)(cur_text_section->data + t);
 119         n = *ptr; /* next value */
 120         *ptr = a - t - 4;
 121         t = n;
 122     }
 123 }
 124
 125 void gsym(int t)
 126 {
 127     gsym_addr(t, ind);
 128 }
 129
 130 /* psym is used to put an instruction with a data field which is a
 131    reference to a symbol. It is in fact the same as oad ! */
 132 #define psym oad
 133
 134 /* instruction + 4 bytes data. Return the address of the data */
 135 static int oad(int c, int s)
 136 {
 137     int ind1;
 138
 139     o(c);
 140     ind1 = ind + 4;
 141     if (ind1 > cur_text_section->data_allocated)
 142         section_realloc(cur_text_section, ind1);
 143     *(int *)(cur_text_section->data + ind) = s;
 144     s = ind;
 145     ind = ind1;
 146     return s;
 147 }
 148
 149 /* output constant with relocation if 'r & VT_SYM' is true */
 150 static void gen_addr32(int r, Sym *sym, int c)
 151 {
 152     if (r & VT_SYM)
 153         greloc(cur_text_section, sym, ind, R_386_32);
 154     gen_le32(c);
 155 }
 156
 157 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 158    opcode bits */
 159 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 160 {
 161     op_reg = op_reg << 3;
 162     if ((r & VT_VALMASK) == VT_CONST) {
 163         /* constant memory reference */
 164         o(0x05 | op_reg);
 165         gen_addr32(r, sym, c);
 166     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 167         /* currently, we use only ebp as base */
 168         if (c == (char)c) {
 169             /* short reference */
 170             o(0x45 | op_reg);
 171             g(c);
 172         } else {
 173             oad(0x85 | op_reg, c);
 174         }
 175     } else {
 176         g(0x00 | op_reg | (r & VT_VALMASK));
 177     }
 178 }
 179
 180
 181 /* load 'r' from value 'sv' */
 182 void load(int r, SValue *sv)
 183 {
 184     int v, t, ft, fc, fr;
 185     SValue v1;
 186
 187     fr = sv->r;
 188     ft = sv->t;
 189     fc = sv->c.ul;
 190
 191     v = fr & VT_VALMASK;
 192     if (fr & VT_LVAL) {
 193         if (v == VT_LLOCAL) {
 194             v1.t = VT_INT;
 195             v1.r = VT_LOCAL | VT_LVAL;
 196             v1.c.ul = fc;
 197             load(r, &v1);
 198             fr = r;
 199         }
 200         if ((ft & VT_BTYPE) == VT_FLOAT) {
 201             o(0xd9); /* flds */
 202             r = 0;
 203         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 204             o(0xdd); /* fldl */
 205             r = 0;
 206         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 207             o(0xdb); /* fldt */
 208             r = 5;
 209         } else if ((ft & VT_TYPE) == VT_BYTE) {
 210             o(0xbe0f);   /* movsbl */
 211         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 212             o(0xb60f);   /* movzbl */
 213         } else if ((ft & VT_TYPE) == VT_SHORT) {
 214             o(0xbf0f);   /* movswl */
 215         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 216             o(0xb70f);   /* movzwl */
 217         } else {
 218             o(0x8b);     /* movl */
 219         }
 220         gen_modrm(r, fr, sv->sym, fc);
 221     } else {
 222         if (v == VT_CONST) {
 223             o(0xb8 + r); /* mov $xx, r */
 224             gen_addr32(fr, sv->sym, fc);
 225         } else if (v == VT_LOCAL) {
 226             o(0x8d); /* lea xxx(%ebp), r */
 227             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 228         } else if (v == VT_CMP) {
 229             oad(0xb8 + r, 0); /* mov $0, r */
 230             o(0x0f); /* setxx %br */
 231             o(fc);
 232             o(0xc0 + r);
 233         } else if (v == VT_JMP || v == VT_JMPI) {
 234             t = v & 1;
 235             oad(0xb8 + r, t); /* mov $1, r */
 236             oad(0xe9, 5); /* jmp after */
 237             gsym(fc);
 238             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 239         } else if (v != r) {
 240             o(0x89);
 241             o(0xc0 + r + v * 8); /* mov v, r */
 242         }
 243     }
 244 }
 245
 246 /* store register 'r' in lvalue 'v' */
 247 void store(int r, SValue *v)
 248 {
 249     int fr, bt, ft, fc;
 250
 251     ft = v->t;
 252     fc = v->c.ul;
 253     fr = v->r & VT_VALMASK;
 254     bt = ft & VT_BTYPE;
 255     /* XXX: incorrect if float reg to reg */
 256     if (bt == VT_FLOAT) {
 257         o(0xd9); /* fsts */
 258         r = 2;
 259     } else if (bt == VT_DOUBLE) {
 260         o(0xdd); /* fstpl */
 261         r = 2;
 262     } else if (bt == VT_LDOUBLE) {
 263         o(0xc0d9); /* fld %st(0) */
 264         o(0xdb); /* fstpt */
 265         r = 7;
 266     } else {
 267         if (bt == VT_SHORT)
 268             o(0x66);
 269         if (bt == VT_BYTE)
 270             o(0x88);
 271         else
 272             o(0x89);
 273     }
 274     if (fr == VT_CONST ||
 275         fr == VT_LOCAL ||
 276         (v->r & VT_LVAL)) {
 277         gen_modrm(r, v->r, v->sym, fc);
 278     } else if (fr != r) {
 279         o(0xc0 + fr + r * 8); /* mov r, fr */
 280     }
 281 }
 282
 283 /* start function call and return function call context */
 284 void gfunc_start(GFuncContext *c, int func_call)
 285 {
 286     c->args_size = 0;
 287     c->func_call = func_call;
 288 }
 289
 290 /* push function parameter which is in (vtop->t, vtop->c). Stack entry
 291    is then popped. */
 292 void gfunc_param(GFuncContext *c)
 293 {
 294     int size, align, r;
 295
 296     if ((vtop->t & VT_BTYPE) == VT_STRUCT) {
 297         size = type_size(vtop->t, &align);
 298         /* align to stack align size */
 299         size = (size + 3) & ~3;
 300         /* allocate the necessary size on stack */
 301         oad(0xec81, size); /* sub $xxx, %esp */
 302         /* generate structure store */
 303         r = get_reg(RC_INT);
 304         o(0x89); /* mov %esp, r */
 305         o(0xe0 + r);
 306         vset(vtop->t, r | VT_LVAL, 0);
 307         vswap();
 308         vstore();
 309         c->args_size += size;
 310     } else if (is_float(vtop->t)) {
 311         gv(RC_FLOAT); /* only one float register */
 312         if ((vtop->t & VT_BTYPE) == VT_FLOAT)
 313             size = 4;
 314         else if ((vtop->t & VT_BTYPE) == VT_DOUBLE)
 315             size = 8;
 316         else
 317             size = 12;
 318         oad(0xec81, size); /* sub $xxx, %esp */
 319         if (size == 12)
 320             o(0x7cdb);
 321         else
 322             o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 323         g(0x24);
 324         g(0x00);
 325         c->args_size += size;
 326     } else {
 327         /* simple type (currently always same size) */
 328         /* XXX: implicit cast ? */
 329         r = gv(RC_INT);
 330         if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 331             size = 8;
 332             o(0x50 + vtop->r2); /* push r */
 333         } else {
 334             size = 4;
 335         }
 336         o(0x50 + r); /* push r */
 337         c->args_size += size;
 338     }
 339     vtop--;
 340 }
 341
 342 static void gadd_sp(int val)
 343 {
 344     if (val == (char)val) {
 345         o(0xc483);
 346         g(val);
 347     } else {
 348         oad(0xc481, val); /* add $xxx, %esp */
 349     }
 350 }
 351
 352 /* generate function call with address in (vtop->t, vtop->c) and free function
 353    context. Stack entry is popped */
 354 void gfunc_call(GFuncContext *c)
 355 {
 356     int r;
 357     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 358         /* constant case */
 359         if (vtop->r & VT_SYM) {
 360             /* relocation case */
 361             greloc(cur_text_section, vtop->sym,
 362                    ind + 1, R_386_PC32);
 363             oad(0xe8, vtop->c.ul - 4);
 364         } else {
 365             oad(0xe8, vtop->c.ul - ind - 5);
 366         }
 367     } else {
 368         /* otherwise, indirect call */
 369         r = gv(RC_INT);
 370         o(0xff); /* call *r */
 371         o(0xd0 + r);
 372     }
 373     if (c->args_size && c->func_call == FUNC_CDECL)
 374         gadd_sp(c->args_size);
 375     vtop--;
 376 }
 377
 378 /* generate function prolog of type 't' */
 379 void gfunc_prolog(int t)
 380 {
 381     int addr, align, size, u, func_call;
 382     Sym *sym;
 383
 384     sym = sym_find((unsigned)t >> VT_STRUCT_SHIFT);
 385     func_call = sym->r;
 386     addr = 8;
 387     /* if the function returns a structure, then add an
 388        implicit pointer parameter */
 389     func_vt = sym->t;
 390     if ((func_vt & VT_BTYPE) == VT_STRUCT) {
 391         func_vc = addr;
 392         addr += 4;
 393     }
 394     /* define parameters */
 395     while ((sym = sym->next) != NULL) {
 396         u = sym->t;
 397         sym_push(sym->v & ~SYM_FIELD, u,
 398                  VT_LOCAL | VT_LVAL, addr);
 399         size = type_size(u, &align);
 400         size = (size + 3) & ~3;
 401 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 402         /* structs are passed as pointer */
 403         if ((u & VT_BTYPE) == VT_STRUCT) {
 404             size = 4;
 405         }
 406 #endif
 407         addr += size;
 408     }
 409     func_ret_sub = 0;
 410     /* pascal type call ? */
 411     if (func_call == FUNC_STDCALL)
 412         func_ret_sub = addr - 8;
 413     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 414     func_sub_sp_offset = oad(0xec81, 0); /* sub $xxx, %esp */
 415     /* leave some room for bound checking code */
 416     if (do_bounds_check) {
 417         oad(0xb8, 0); /* lbound section pointer */
 418         oad(0xb8, 0); /* call to function */
 419         func_bound_offset = lbounds_section->data_offset;
 420     }
 421 }
 422
 423 /* generate function epilog */
 424 void gfunc_epilog(void)
 425 {
 426 #ifdef CONFIG_TCC_BCHECK
 427     if (do_bounds_check && func_bound_offset != lbounds_section->data_offset) {
 428         int saved_ind;
 429         int *bounds_ptr;
 430         Sym *sym, *sym_data;
 431         /* add end of table info */
 432         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 433         *bounds_ptr = 0;
 434         /* generate bound local allocation */
 435         saved_ind = ind;
 436         ind = func_sub_sp_offset + 4;
 437         sym_data = get_sym_ref(char_pointer_type, lbounds_section,
 438                                func_bound_offset, lbounds_section->data_offset);
 439         greloc(cur_text_section, sym_data,
 440                ind + 1, R_386_32);
 441         oad(0xb8, 0); /* mov %eax, xxx */
 442         sym = external_global_sym(TOK___bound_local_new, func_old_type, 0);
 443         greloc(cur_text_section, sym,
 444                ind + 1, R_386_PC32);
 445         oad(0xe8, -4);
 446         ind = saved_ind;
 447         /* generate bound check local freeing */
 448         o(0x5250); /* save returned value, if any */
 449         greloc(cur_text_section, sym_data,
 450                ind + 1, R_386_32);
 451         oad(0xb8, 0); /* mov %eax, xxx */
 452         sym = external_global_sym(TOK___bound_local_delete, func_old_type, 0);
 453         greloc(cur_text_section, sym,
 454                ind + 1, R_386_PC32);
 455         oad(0xe8, -4);
 456         o(0x585a); /* restore returned value, if any */
 457     }
 458 #endif
 459     o(0xc9); /* leave */
 460     if (func_ret_sub == 0) {
 461         o(0xc3); /* ret */
 462     } else {
 463         o(0xc2); /* ret n */
 464         g(func_ret_sub);
 465         g(func_ret_sub >> 8);
 466     }
 467     /* align local size to word & save local variables */
 468     *(int *)(cur_text_section->data + func_sub_sp_offset) = (-loc + 3) & -4;
 469 }
 470
 471 /* generate a jump to a label */
 472 int gjmp(int t)
 473 {
 474     return psym(0xe9, t);
 475 }
 476
 477 /* generate a jump to a fixed address */
 478 void gjmp_addr(int a)
 479 {
 480     oad(0xe9, a - ind - 5);
 481 }
 482
 483 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 484 int gtst(int inv, int t)
 485 {
 486     int v, *p;
 487     v = vtop->r & VT_VALMASK;
 488     if (v == VT_CMP) {
 489         /* fast case : can jump directly since flags are set */
 490         g(0x0f);
 491         t = psym((vtop->c.i - 16) ^ inv, t);
 492     } else if (v == VT_JMP || v == VT_JMPI) {
 493         /* && or || optimization */
 494         if ((v & 1) == inv) {
 495             /* insert vtop->c jump list in t */
 496             p = &vtop->c.i;
 497             while (*p != 0)
 498                 p = (int *)(cur_text_section->data + *p);
 499             *p = t;
 500             t = vtop->c.i;
 501         } else {
 502             t = gjmp(t);
 503             gsym(vtop->c.i);
 504         }
 505     } else {
 506         if (is_float(vtop->t)) {
 507             vpushi(0);
 508             gen_op(TOK_NE);
 509         }
 510         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 511             /* constant jmp optimization */
 512             if ((vtop->c.i != 0) != inv)
 513                 t = gjmp(t);
 514         } else {
 515             v = gv(RC_INT);
 516             o(0x85);
 517             o(0xc0 + v * 9);
 518             g(0x0f);
 519             t = psym(0x85 ^ inv, t);
 520         }
 521     }
 522     vtop--;
 523     return t;
 524 }
 525
 526 /* generate an integer binary operation */
 527 void gen_opi(int op)
 528 {
 529     int r, fr, opc, c;
 530
 531     switch(op) {
 532     case '+':
 533     case TOK_ADDC1: /* add with carry generation */
 534         opc = 0;
 535     gen_op8:
 536         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 537             /* constant case */
 538             vswap();
 539             r = gv(RC_INT);
 540             vswap();
 541             c = vtop->c.i;
 542             if (c == (char)c) {
 543                 /* XXX: generate inc and dec for smaller code ? */
 544                 o(0x83);
 545                 o(0xc0 | (opc << 3) | r);
 546                 g(c);
 547             } else {
 548                 o(0x81);
 549                 oad(0xc0 | (opc << 3) | r, c);
 550             }
 551         } else {
 552             gv2(RC_INT, RC_INT);
 553             r = vtop[-1].r;
 554             fr = vtop[0].r;
 555             o((opc << 3) | 0x01);
 556             o(0xc0 + r + fr * 8);
 557         }
 558         vtop--;
 559         if (op >= TOK_ULT && op <= TOK_GT) {
 560             vtop--;
 561             vset(VT_INT, VT_CMP, op);
 562         }
 563         break;
 564     case '-':
 565     case TOK_SUBC1: /* sub with carry generation */
 566         opc = 5;
 567         goto gen_op8;
 568     case TOK_ADDC2: /* add with carry use */
 569         opc = 2;
 570         goto gen_op8;
 571     case TOK_SUBC2: /* sub with carry use */
 572         opc = 3;
 573         goto gen_op8;
 574     case '&':
 575         opc = 4;
 576         goto gen_op8;
 577     case '^':
 578         opc = 6;
 579         goto gen_op8;
 580     case '|':
 581         opc = 1;
 582         goto gen_op8;
 583     case '*':
 584         gv2(RC_INT, RC_INT);
 585         r = vtop[-1].r;
 586         fr = vtop[0].r;
 587         vtop--;
 588         o(0xaf0f); /* imul fr, r */
 589         o(0xc0 + fr + r * 8);
 590         break;
 591     case TOK_SHL:
 592         opc = 4;
 593         goto gen_shift;
 594     case TOK_SHR:
 595         opc = 5;
 596         goto gen_shift;
 597     case TOK_SAR:
 598         opc = 7;
 599     gen_shift:
 600         opc = 0xc0 | (opc << 3);
 601         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 602             /* constant case */
 603             vswap();
 604             r = gv(RC_INT);
 605             vswap();
 606             c = vtop->c.i & 0x1f;
 607             o(0xc1); /* shl/shr/sar $xxx, r */
 608             o(opc | r);
 609             g(c);
 610         } else {
 611             /* we generate the shift in ecx */
 612             gv2(RC_INT, RC_ECX);
 613             r = vtop[-1].r;
 614             o(0xd3); /* shl/shr/sar %cl, r */
 615             o(opc | r);
 616         }
 617         vtop--;
 618         break;
 619     case '/':
 620     case TOK_UDIV:
 621     case TOK_PDIV:
 622     case '%':
 623     case TOK_UMOD:
 624     case TOK_UMULL:
 625         /* first operand must be in eax */
 626         /* XXX: need better constraint for second operand */
 627         gv2(RC_EAX, RC_ECX);
 628         r = vtop[-1].r;
 629         fr = vtop[0].r;
 630         vtop--;
 631         save_reg(REG_EDX);
 632         if (op == TOK_UMULL) {
 633             o(0xf7); /* mul fr */
 634             o(0xe0 + fr);
 635             vtop->r2 = REG_EDX;
 636             r = REG_EAX;
 637         } else {
 638             if (op == TOK_UDIV || op == TOK_UMOD) {
 639                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 640                 o(0xf0 + fr);
 641             } else {
 642                 o(0xf799); /* cltd, idiv fr, %eax */
 643                 o(0xf8 + fr);
 644             }
 645             if (op == '%' || op == TOK_UMOD)
 646                 r = REG_EDX;
 647             else
 648                 r = REG_EAX;
 649         }
 650         vtop->r = r;
 651         break;
 652     default:
 653         opc = 7;
 654         goto gen_op8;
 655     }
 656 }
 657
 658 /* generate a floating point operation 'v = t1 op t2' instruction. The
 659    two operands are guaranted to have the same floating point type */
 660 /* XXX: need to use ST1 too */
 661 void gen_opf(int op)
 662 {
 663     int a, ft, fc, swapped, r;
 664
 665     /* convert constants to memory references */
 666     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 667         vswap();
 668         gv(RC_FLOAT);
 669         vswap();
 670     }
 671     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 672         gv(RC_FLOAT);
 673
 674     /* must put at least one value in the floating point register */
 675     if ((vtop[-1].r & VT_LVAL) &&
 676         (vtop[0].r & VT_LVAL)) {
 677         vswap();
 678         gv(RC_FLOAT);
 679         vswap();
 680     }
 681     swapped = 0;
 682     /* swap the stack if needed so that t1 is the register and t2 is
 683        the memory reference */
 684     if (vtop[-1].r & VT_LVAL) {
 685         vswap();
 686         swapped = 1;
 687     }
 688     if (op >= TOK_ULT && op <= TOK_GT) {
 689         /* load on stack second operand */
 690         load(REG_ST0, vtop);
 691         save_reg(REG_EAX); /* eax is used by FP comparison code */
 692         if (op == TOK_GE || op == TOK_GT)
 693             swapped = !swapped;
 694         else if (op == TOK_EQ || op == TOK_NE)
 695             swapped = 0;
 696         if (swapped)
 697             o(0xc9d9); /* fxch %st(1) */
 698         o(0xe9da); /* fucompp */
 699         o(0xe0df); /* fnstsw %ax */
 700         if (op == TOK_EQ) {
 701             o(0x45e480); /* and $0x45, %ah */
 702             o(0x40fC80); /* cmp $0x40, %ah */
 703         } else if (op == TOK_NE) {
 704             o(0x45e480); /* and $0x45, %ah */
 705             o(0x40f480); /* xor $0x40, %ah */
 706             op = TOK_NE;
 707         } else if (op == TOK_GE || op == TOK_LE) {
 708             o(0x05c4f6); /* test $0x05, %ah */
 709             op = TOK_EQ;
 710         } else {
 711             o(0x45c4f6); /* test $0x45, %ah */
 712             op = TOK_EQ;
 713         }
 714         vtop--;
 715         vtop->r = VT_CMP;
 716         vtop->c.i = op;
 717     } else {
 718         /* no memory reference possible for long double operations */
 719         if ((vtop->t & VT_BTYPE) == VT_LDOUBLE) {
 720             load(REG_ST0, vtop);
 721             swapped = !swapped;
 722         }
 723
 724         switch(op) {
 725         default:
 726         case '+':
 727             a = 0;
 728             break;
 729         case '-':
 730             a = 4;
 731             if (swapped)
 732                 a++;
 733             break;
 734         case '*':
 735             a = 1;
 736             break;
 737         case '/':
 738             a = 6;
 739             if (swapped)
 740                 a++;
 741             break;
 742         }
 743         ft = vtop->t;
 744         fc = vtop->c.ul;
 745         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 746             o(0xde); /* fxxxp %st, %st(1) */
 747             o(0xc1 + (a << 3));
 748         } else {
 749             /* if saved lvalue, then we must reload it */
 750             r = vtop->r;
 751             if ((r & VT_VALMASK) == VT_LLOCAL) {
 752                 SValue v1;
 753                 r = get_reg(RC_INT);
 754                 v1.t = VT_INT;
 755                 v1.r = VT_LOCAL | VT_LVAL;
 756                 v1.c.ul = fc;
 757                 load(r, &v1);
 758                 fc = 0;
 759             }
 760
 761             if ((ft & VT_BTYPE) == VT_DOUBLE)
 762                 o(0xdc);
 763             else
 764                 o(0xd8);
 765             gen_modrm(a, r, vtop->sym, fc);
 766         }
 767         vtop--;
 768     }
 769 }
 770
 771 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 772    and 'long long' cases. */
 773 void gen_cvt_itof(int t)
 774 {
 775     save_reg(REG_ST0);
 776     gv(RC_INT);
 777     if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 778         /* signed long long to float/double/long double (unsigned case
 779            is handled generically) */
 780         o(0x50 + vtop->r2); /* push r2 */
 781         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 782         o(0x242cdf); /* fildll (%esp) */
 783         o(0x08c483); /* add $8, %esp */
 784     } else if ((vtop->t & (VT_BTYPE | VT_UNSIGNED)) ==
 785                (VT_INT | VT_UNSIGNED)) {
 786         /* unsigned int to float/double/long double */
 787         o(0x6a); /* push $0 */
 788         g(0x00);
 789         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 790         o(0x242cdf); /* fildll (%esp) */
 791         o(0x08c483); /* add $8, %esp */
 792     } else {
 793         /* int to float/double/long double */
 794         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 795         o(0x2404db); /* fildl (%esp) */
 796         o(0x04c483); /* add $4, %esp */
 797     }
 798     vtop->r = REG_ST0;
 799 }
 800
 801 /* convert fp to int 't' type */
 802 /* XXX: handle long long case */
 803 void gen_cvt_ftoi(int t)
 804 {
 805     int r, r2, size;
 806     Sym *sym;
 807
 808     gv(RC_FLOAT);
 809     if (t != VT_INT)
 810         size = 8;
 811     else
 812         size = 4;
 813
 814     o(0x2dd9); /* ldcw xxx */
 815     sym = external_global_sym(TOK___tcc_int_fpu_control,
 816                               VT_SHORT | VT_UNSIGNED, VT_LVAL);
 817     greloc(cur_text_section, sym,
 818            ind, R_386_32);
 819     gen_le32(0);
 820
 821     oad(0xec81, size); /* sub $xxx, %esp */
 822     if (size == 4)
 823         o(0x1cdb); /* fistpl */
 824     else
 825         o(0x3cdf); /* fistpll */
 826     o(0x24);
 827     o(0x2dd9); /* ldcw xxx */
 828     sym = external_global_sym(TOK___tcc_fpu_control,
 829                               VT_SHORT | VT_UNSIGNED, VT_LVAL);
 830     greloc(cur_text_section, sym,
 831            ind, R_386_32);
 832     gen_le32(0);
 833
 834     r = get_reg(RC_INT);
 835     o(0x58 + r); /* pop r */
 836     if (size == 8) {
 837         if (t == VT_LLONG) {
 838             vtop->r = r; /* mark reg as used */
 839             r2 = get_reg(RC_INT);
 840             o(0x58 + r2); /* pop r2 */
 841             vtop->r2 = r2;
 842         } else {
 843             o(0x04c483); /* add $4, %esp */
 844         }
 845     }
 846     vtop->r = r;
 847 }
 848
 849 /* convert from one floating point type to another */
 850 void gen_cvt_ftof(int t)
 851 {
 852     /* all we have to do on i386 is to put the float in a register */
 853     gv(RC_FLOAT);
 854 }
 855
 856 /* bound check support functions */
 857 #ifdef CONFIG_TCC_BCHECK
 858
 859 /* generate a bounded pointer addition */
 860 void gen_bounded_ptr_add(void)
 861 {
 862     Sym *sym;
 863
 864     /* prepare fast i386 function call (args in eax and edx) */
 865     gv2(RC_EAX, RC_EDX);
 866     /* save all temporary registers */
 867     vtop -= 2;
 868     save_regs(0);
 869     /* do a fast function call */
 870     sym = external_global_sym(TOK___bound_ptr_add, func_old_type, 0);
 871     greloc(cur_text_section, sym,
 872            ind + 1, R_386_PC32);
 873     oad(0xe8, -4);
 874     /* returned pointer is in eax */
 875     vtop++;
 876     vtop->r = REG_EAX | VT_BOUNDED;
 877     /* address of bounding function call point */
 878     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 879 }
 880
 881 /* patch pointer addition in vtop so that pointer dereferencing is
 882    also tested */
 883 void gen_bounded_ptr_deref(void)
 884 {
 885     int func;
 886     int size, align;
 887     Elf32_Rel *rel;
 888     Sym *sym;
 889
 890     size = 0;
 891     /* XXX: put that code in generic part of tcc */
 892     if (!is_float(vtop->t)) {
 893         if (vtop->r & VT_LVAL_BYTE)
 894             size = 1;
 895         else if (vtop->r & VT_LVAL_SHORT)
 896             size = 2;
 897     }
 898     if (!size)
 899         size = type_size(vtop->t, &align);
 900     switch(size) {
 901     case  1: func = TOK___bound_ptr_indir1; break;
 902     case  2: func = TOK___bound_ptr_indir2; break;
 903     case  4: func = TOK___bound_ptr_indir4; break;
 904     case  8: func = TOK___bound_ptr_indir8; break;
 905     case 12: func = TOK___bound_ptr_indir12; break;
 906     case 16: func = TOK___bound_ptr_indir16; break;
 907     default:
 908         error("unhandled size when derefencing bounded pointer");
 909         func = 0;
 910         break;
 911     }
 912
 913     /* patch relocation */
 914     /* XXX: find a better solution ? */
 915     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
 916     sym = external_global_sym(func, func_old_type, 0);
 917     if (!sym->c)
 918         put_extern_sym(sym, NULL, 0, 0);
 919     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
 920 }
 921 #endif
 922
 923 /* end of X86 code generator */
 924 /*************************************************************/
 925