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             o(0x05eb); /* 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         } else {
 364             /* put an empty PC32 relocation */
 365             put_elf_reloc(symtab_section, cur_text_section,
 366                           ind + 1, R_386_PC32, 0);
 367         }
 368         oad(0xe8, vtop->c.ul - 4);
 369     } else {
 370         /* otherwise, indirect call */
 371         r = gv(RC_INT);
 372         o(0xff); /* call *r */
 373         o(0xd0 + r);
 374     }
 375     if (c->args_size && c->func_call == FUNC_CDECL)
 376         gadd_sp(c->args_size);
 377     vtop--;
 378 }
 379
 380 /* generate function prolog of type 't' */
 381 void gfunc_prolog(int t)
 382 {
 383     int addr, align, size, u, func_call;
 384     Sym *sym;
 385
 386     sym = sym_find((unsigned)t >> VT_STRUCT_SHIFT);
 387     func_call = sym->r;
 388     addr = 8;
 389     /* if the function returns a structure, then add an
 390        implicit pointer parameter */
 391     func_vt = sym->t;
 392     if ((func_vt & VT_BTYPE) == VT_STRUCT) {
 393         func_vc = addr;
 394         addr += 4;
 395     }
 396     /* define parameters */
 397     while ((sym = sym->next) != NULL) {
 398         u = sym->t;
 399         sym_push(sym->v & ~SYM_FIELD, u,
 400                  VT_LOCAL | VT_LVAL, addr);
 401         size = type_size(u, &align);
 402         size = (size + 3) & ~3;
 403 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 404         /* structs are passed as pointer */
 405         if ((u & VT_BTYPE) == VT_STRUCT) {
 406             size = 4;
 407         }
 408 #endif
 409         addr += size;
 410     }
 411     func_ret_sub = 0;
 412     /* pascal type call ? */
 413     if (func_call == FUNC_STDCALL)
 414         func_ret_sub = addr - 8;
 415     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 416     func_sub_sp_offset = oad(0xec81, 0); /* sub $xxx, %esp */
 417     /* leave some room for bound checking code */
 418     if (do_bounds_check) {
 419         oad(0xb8, 0); /* lbound section pointer */
 420         oad(0xb8, 0); /* call to function */
 421         func_bound_offset = lbounds_section->data_offset;
 422     }
 423 }
 424
 425 /* generate function epilog */
 426 void gfunc_epilog(void)
 427 {
 428 #ifdef CONFIG_TCC_BCHECK
 429     if (do_bounds_check && func_bound_offset != lbounds_section->data_offset) {
 430         int saved_ind;
 431         int *bounds_ptr;
 432         Sym *sym, *sym_data;
 433         /* add end of table info */
 434         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 435         *bounds_ptr = 0;
 436         /* generate bound local allocation */
 437         saved_ind = ind;
 438         ind = func_sub_sp_offset + 4;
 439         sym_data = get_sym_ref(char_pointer_type, lbounds_section,
 440                                func_bound_offset, lbounds_section->data_offset);
 441         greloc(cur_text_section, sym_data,
 442                ind + 1, R_386_32);
 443         oad(0xb8, 0); /* mov %eax, xxx */
 444         sym = external_global_sym(TOK___bound_local_new, func_old_type, 0);
 445         greloc(cur_text_section, sym,
 446                ind + 1, R_386_PC32);
 447         oad(0xe8, -4);
 448         ind = saved_ind;
 449         /* generate bound check local freeing */
 450         o(0x5250); /* save returned value, if any */
 451         greloc(cur_text_section, sym_data,
 452                ind + 1, R_386_32);
 453         oad(0xb8, 0); /* mov %eax, xxx */
 454         sym = external_global_sym(TOK___bound_local_delete, func_old_type, 0);
 455         greloc(cur_text_section, sym,
 456                ind + 1, R_386_PC32);
 457         oad(0xe8, -4);
 458         o(0x585a); /* restore returned value, if any */
 459     }
 460 #endif
 461     o(0xc9); /* leave */
 462     if (func_ret_sub == 0) {
 463         o(0xc3); /* ret */
 464     } else {
 465         o(0xc2); /* ret n */
 466         g(func_ret_sub);
 467         g(func_ret_sub >> 8);
 468     }
 469     /* align local size to word & save local variables */
 470     *(int *)(cur_text_section->data + func_sub_sp_offset) = (-loc + 3) & -4;
 471 }
 472
 473 /* generate a jump to a label */
 474 int gjmp(int t)
 475 {
 476     return psym(0xe9, t);
 477 }
 478
 479 /* generate a jump to a fixed address */
 480 void gjmp_addr(int a)
 481 {
 482     int r;
 483     r = a - ind - 2;
 484     if (r == (char)r) {
 485         g(0xeb);
 486         g(r);
 487     } else {
 488         oad(0xe9, a - ind - 5);
 489     }
 490 }
 491
 492 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 493 int gtst(int inv, int t)
 494 {
 495     int v, *p;
 496     v = vtop->r & VT_VALMASK;
 497     if (v == VT_CMP) {
 498         /* fast case : can jump directly since flags are set */
 499         g(0x0f);
 500         t = psym((vtop->c.i - 16) ^ inv, t);
 501     } else if (v == VT_JMP || v == VT_JMPI) {
 502         /* && or || optimization */
 503         if ((v & 1) == inv) {
 504             /* insert vtop->c jump list in t */
 505             p = &vtop->c.i;
 506             while (*p != 0)
 507                 p = (int *)(cur_text_section->data + *p);
 508             *p = t;
 509             t = vtop->c.i;
 510         } else {
 511             t = gjmp(t);
 512             gsym(vtop->c.i);
 513         }
 514     } else {
 515         if (is_float(vtop->t)) {
 516             vpushi(0);
 517             gen_op(TOK_NE);
 518         }
 519         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 520             /* constant jmp optimization */
 521             if ((vtop->c.i != 0) != inv)
 522                 t = gjmp(t);
 523         } else {
 524             v = gv(RC_INT);
 525             o(0x85);
 526             o(0xc0 + v * 9);
 527             g(0x0f);
 528             t = psym(0x85 ^ inv, t);
 529         }
 530     }
 531     vtop--;
 532     return t;
 533 }
 534
 535 /* generate an integer binary operation */
 536 void gen_opi(int op)
 537 {
 538     int r, fr, opc, c;
 539
 540     switch(op) {
 541     case '+':
 542     case TOK_ADDC1: /* add with carry generation */
 543         opc = 0;
 544     gen_op8:
 545         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 546             /* constant case */
 547             vswap();
 548             r = gv(RC_INT);
 549             vswap();
 550             c = vtop->c.i;
 551             if (c == (char)c) {
 552                 /* XXX: generate inc and dec for smaller code ? */
 553                 o(0x83);
 554                 o(0xc0 | (opc << 3) | r);
 555                 g(c);
 556             } else {
 557                 o(0x81);
 558                 oad(0xc0 | (opc << 3) | r, c);
 559             }
 560         } else {
 561             gv2(RC_INT, RC_INT);
 562             r = vtop[-1].r;
 563             fr = vtop[0].r;
 564             o((opc << 3) | 0x01);
 565             o(0xc0 + r + fr * 8);
 566         }
 567         vtop--;
 568         if (op >= TOK_ULT && op <= TOK_GT) {
 569             vtop--;
 570             vset(VT_INT, VT_CMP, op);
 571         }
 572         break;
 573     case '-':
 574     case TOK_SUBC1: /* sub with carry generation */
 575         opc = 5;
 576         goto gen_op8;
 577     case TOK_ADDC2: /* add with carry use */
 578         opc = 2;
 579         goto gen_op8;
 580     case TOK_SUBC2: /* sub with carry use */
 581         opc = 3;
 582         goto gen_op8;
 583     case '&':
 584         opc = 4;
 585         goto gen_op8;
 586     case '^':
 587         opc = 6;
 588         goto gen_op8;
 589     case '|':
 590         opc = 1;
 591         goto gen_op8;
 592     case '*':
 593         gv2(RC_INT, RC_INT);
 594         r = vtop[-1].r;
 595         fr = vtop[0].r;
 596         vtop--;
 597         o(0xaf0f); /* imul fr, r */
 598         o(0xc0 + fr + r * 8);
 599         break;
 600     case TOK_SHL:
 601         opc = 4;
 602         goto gen_shift;
 603     case TOK_SHR:
 604         opc = 5;
 605         goto gen_shift;
 606     case TOK_SAR:
 607         opc = 7;
 608     gen_shift:
 609         opc = 0xc0 | (opc << 3);
 610         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 611             /* constant case */
 612             vswap();
 613             r = gv(RC_INT);
 614             vswap();
 615             c = vtop->c.i & 0x1f;
 616             o(0xc1); /* shl/shr/sar $xxx, r */
 617             o(opc | r);
 618             g(c);
 619         } else {
 620             /* we generate the shift in ecx */
 621             gv2(RC_INT, RC_ECX);
 622             r = vtop[-1].r;
 623             o(0xd3); /* shl/shr/sar %cl, r */
 624             o(opc | r);
 625         }
 626         vtop--;
 627         break;
 628     case '/':
 629     case TOK_UDIV:
 630     case TOK_PDIV:
 631     case '%':
 632     case TOK_UMOD:
 633     case TOK_UMULL:
 634         /* first operand must be in eax */
 635         /* XXX: need better constraint for second operand */
 636         gv2(RC_EAX, RC_ECX);
 637         r = vtop[-1].r;
 638         fr = vtop[0].r;
 639         vtop--;
 640         save_reg(REG_EDX);
 641         if (op == TOK_UMULL) {
 642             o(0xf7); /* mul fr */
 643             o(0xe0 + fr);
 644             vtop->r2 = REG_EDX;
 645             r = REG_EAX;
 646         } else {
 647             if (op == TOK_UDIV || op == TOK_UMOD) {
 648                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 649                 o(0xf0 + fr);
 650             } else {
 651                 o(0xf799); /* cltd, idiv fr, %eax */
 652                 o(0xf8 + fr);
 653             }
 654             if (op == '%' || op == TOK_UMOD)
 655                 r = REG_EDX;
 656             else
 657                 r = REG_EAX;
 658         }
 659         vtop->r = r;
 660         break;
 661     default:
 662         opc = 7;
 663         goto gen_op8;
 664     }
 665 }
 666
 667 /* generate a floating point operation 'v = t1 op t2' instruction. The
 668    two operands are guaranted to have the same floating point type */
 669 /* XXX: need to use ST1 too */
 670 void gen_opf(int op)
 671 {
 672     int a, ft, fc, swapped, r;
 673
 674     /* convert constants to memory references */
 675     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 676         vswap();
 677         gv(RC_FLOAT);
 678         vswap();
 679     }
 680     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 681         gv(RC_FLOAT);
 682
 683     /* must put at least one value in the floating point register */
 684     if ((vtop[-1].r & VT_LVAL) &&
 685         (vtop[0].r & VT_LVAL)) {
 686         vswap();
 687         gv(RC_FLOAT);
 688         vswap();
 689     }
 690     swapped = 0;
 691     /* swap the stack if needed so that t1 is the register and t2 is
 692        the memory reference */
 693     if (vtop[-1].r & VT_LVAL) {
 694         vswap();
 695         swapped = 1;
 696     }
 697     if (op >= TOK_ULT && op <= TOK_GT) {
 698         /* load on stack second operand */
 699         load(REG_ST0, vtop);
 700         save_reg(REG_EAX); /* eax is used by FP comparison code */
 701         if (op == TOK_GE || op == TOK_GT)
 702             swapped = !swapped;
 703         else if (op == TOK_EQ || op == TOK_NE)
 704             swapped = 0;
 705         if (swapped)
 706             o(0xc9d9); /* fxch %st(1) */
 707         o(0xe9da); /* fucompp */
 708         o(0xe0df); /* fnstsw %ax */
 709         if (op == TOK_EQ) {
 710             o(0x45e480); /* and $0x45, %ah */
 711             o(0x40fC80); /* cmp $0x40, %ah */
 712         } else if (op == TOK_NE) {
 713             o(0x45e480); /* and $0x45, %ah */
 714             o(0x40f480); /* xor $0x40, %ah */
 715             op = TOK_NE;
 716         } else if (op == TOK_GE || op == TOK_LE) {
 717             o(0x05c4f6); /* test $0x05, %ah */
 718             op = TOK_EQ;
 719         } else {
 720             o(0x45c4f6); /* test $0x45, %ah */
 721             op = TOK_EQ;
 722         }
 723         vtop--;
 724         vtop->r = VT_CMP;
 725         vtop->c.i = op;
 726     } else {
 727         /* no memory reference possible for long double operations */
 728         if ((vtop->t & VT_BTYPE) == VT_LDOUBLE) {
 729             load(REG_ST0, vtop);
 730             swapped = !swapped;
 731         }
 732
 733         switch(op) {
 734         default:
 735         case '+':
 736             a = 0;
 737             break;
 738         case '-':
 739             a = 4;
 740             if (swapped)
 741                 a++;
 742             break;
 743         case '*':
 744             a = 1;
 745             break;
 746         case '/':
 747             a = 6;
 748             if (swapped)
 749                 a++;
 750             break;
 751         }
 752         ft = vtop->t;
 753         fc = vtop->c.ul;
 754         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 755             o(0xde); /* fxxxp %st, %st(1) */
 756             o(0xc1 + (a << 3));
 757         } else {
 758             /* if saved lvalue, then we must reload it */
 759             r = vtop->r;
 760             if ((r & VT_VALMASK) == VT_LLOCAL) {
 761                 SValue v1;
 762                 r = get_reg(RC_INT);
 763                 v1.t = VT_INT;
 764                 v1.r = VT_LOCAL | VT_LVAL;
 765                 v1.c.ul = fc;
 766                 load(r, &v1);
 767                 fc = 0;
 768             }
 769
 770             if ((ft & VT_BTYPE) == VT_DOUBLE)
 771                 o(0xdc);
 772             else
 773                 o(0xd8);
 774             gen_modrm(a, r, vtop->sym, fc);
 775         }
 776         vtop--;
 777     }
 778 }
 779
 780 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 781    and 'long long' cases. */
 782 void gen_cvt_itof(int t)
 783 {
 784     save_reg(REG_ST0);
 785     gv(RC_INT);
 786     if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 787         /* signed long long to float/double/long double (unsigned case
 788            is handled generically) */
 789         o(0x50 + vtop->r2); /* push r2 */
 790         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 791         o(0x242cdf); /* fildll (%esp) */
 792         o(0x08c483); /* add $8, %esp */
 793     } else if ((vtop->t & (VT_BTYPE | VT_UNSIGNED)) ==
 794                (VT_INT | VT_UNSIGNED)) {
 795         /* unsigned int to float/double/long double */
 796         o(0x6a); /* push $0 */
 797         g(0x00);
 798         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 799         o(0x242cdf); /* fildll (%esp) */
 800         o(0x08c483); /* add $8, %esp */
 801     } else {
 802         /* int to float/double/long double */
 803         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 804         o(0x2404db); /* fildl (%esp) */
 805         o(0x04c483); /* add $4, %esp */
 806     }
 807     vtop->r = REG_ST0;
 808 }
 809
 810 /* convert fp to int 't' type */
 811 /* XXX: handle long long case */
 812 void gen_cvt_ftoi(int t)
 813 {
 814     int r, r2, size;
 815     Sym *sym;
 816
 817     gv(RC_FLOAT);
 818     if (t != VT_INT)
 819         size = 8;
 820     else
 821         size = 4;
 822
 823     o(0x2dd9); /* ldcw xxx */
 824     sym = external_global_sym(TOK___tcc_int_fpu_control,
 825                               VT_SHORT | VT_UNSIGNED, VT_LVAL);
 826     greloc(cur_text_section, sym,
 827            ind, R_386_32);
 828     gen_le32(0);
 829
 830     oad(0xec81, size); /* sub $xxx, %esp */
 831     if (size == 4)
 832         o(0x1cdb); /* fistpl */
 833     else
 834         o(0x3cdf); /* fistpll */
 835     o(0x24);
 836     o(0x2dd9); /* ldcw xxx */
 837     sym = external_global_sym(TOK___tcc_fpu_control,
 838                               VT_SHORT | VT_UNSIGNED, VT_LVAL);
 839     greloc(cur_text_section, sym,
 840            ind, R_386_32);
 841     gen_le32(0);
 842
 843     r = get_reg(RC_INT);
 844     o(0x58 + r); /* pop r */
 845     if (size == 8) {
 846         if (t == VT_LLONG) {
 847             vtop->r = r; /* mark reg as used */
 848             r2 = get_reg(RC_INT);
 849             o(0x58 + r2); /* pop r2 */
 850             vtop->r2 = r2;
 851         } else {
 852             o(0x04c483); /* add $4, %esp */
 853         }
 854     }
 855     vtop->r = r;
 856 }
 857
 858 /* convert from one floating point type to another */
 859 void gen_cvt_ftof(int t)
 860 {
 861     /* all we have to do on i386 is to put the float in a register */
 862     gv(RC_FLOAT);
 863 }
 864
 865 /* bound check support functions */
 866 #ifdef CONFIG_TCC_BCHECK
 867
 868 /* generate a bounded pointer addition */
 869 void gen_bounded_ptr_add(void)
 870 {
 871     Sym *sym;
 872
 873     /* prepare fast i386 function call (args in eax and edx) */
 874     gv2(RC_EAX, RC_EDX);
 875     /* save all temporary registers */
 876     vtop -= 2;
 877     save_regs(0);
 878     /* do a fast function call */
 879     sym = external_global_sym(TOK___bound_ptr_add, func_old_type, 0);
 880     greloc(cur_text_section, sym,
 881            ind + 1, R_386_PC32);
 882     oad(0xe8, -4);
 883     /* returned pointer is in eax */
 884     vtop++;
 885     vtop->r = REG_EAX | VT_BOUNDED;
 886     /* address of bounding function call point */
 887     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 888 }
 889
 890 /* patch pointer addition in vtop so that pointer dereferencing is
 891    also tested */
 892 void gen_bounded_ptr_deref(void)
 893 {
 894     int func;
 895     int size, align;
 896     Elf32_Rel *rel;
 897     Sym *sym;
 898
 899     size = 0;
 900     /* XXX: put that code in generic part of tcc */
 901     if (!is_float(vtop->t)) {
 902         if (vtop->r & VT_LVAL_BYTE)
 903             size = 1;
 904         else if (vtop->r & VT_LVAL_SHORT)
 905             size = 2;
 906     }
 907     if (!size)
 908         size = type_size(vtop->t, &align);
 909     switch(size) {
 910     case  1: func = TOK___bound_ptr_indir1; break;
 911     case  2: func = TOK___bound_ptr_indir2; break;
 912     case  4: func = TOK___bound_ptr_indir4; break;
 913     case  8: func = TOK___bound_ptr_indir8; break;
 914     case 12: func = TOK___bound_ptr_indir12; break;
 915     case 16: func = TOK___bound_ptr_indir16; break;
 916     default:
 917         error("unhandled size when derefencing bounded pointer");
 918         func = 0;
 919         break;
 920     }
 921
 922     /* patch relocation */
 923     /* XXX: find a better solution ? */
 924     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
 925     sym = external_global_sym(func, func_old_type, 0);
 926     if (!sym->c)
 927         put_extern_sym(sym, NULL, 0, 0);
 928     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
 929 }
 930 #endif
 931
 932 /* end of X86 code generator */
 933 /*************************************************************/
 934