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 /* 'is_jmp' is '1' if it is a jump */
 353 static void gcall_or_jmp(int is_jmp)
 354 {
 355     int r;
 356     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 357         /* constant case */
 358         if (vtop->r & VT_SYM) {
 359             /* relocation case */
 360             greloc(cur_text_section, vtop->sym,
 361                    ind + 1, R_386_PC32);
 362         } else {
 363             /* put an empty PC32 relocation */
 364             put_elf_reloc(symtab_section, cur_text_section,
 365                           ind + 1, R_386_PC32, 0);
 366         }
 367         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 368     } else {
 369         /* otherwise, indirect call */
 370         r = gv(RC_INT);
 371         o(0xff); /* call/jmp *r */
 372         o(0xd0 + r + (is_jmp << 4));
 373     }
 374 }
 375
 376 /* generate function call with address in (vtop->t, vtop->c) and free function
 377    context. Stack entry is popped */
 378 void gfunc_call(GFuncContext *c)
 379 {
 380     gcall_or_jmp(0);
 381     if (c->args_size && c->func_call == FUNC_CDECL)
 382         gadd_sp(c->args_size);
 383     vtop--;
 384 }
 385
 386 /* generate function prolog of type 't' */
 387 void gfunc_prolog(int t)
 388 {
 389     int addr, align, size, u, func_call;
 390     Sym *sym;
 391
 392     sym = sym_find((unsigned)t >> VT_STRUCT_SHIFT);
 393     func_call = sym->r;
 394     addr = 8;
 395     /* if the function returns a structure, then add an
 396        implicit pointer parameter */
 397     func_vt = sym->t;
 398     if ((func_vt & VT_BTYPE) == VT_STRUCT) {
 399         func_vc = addr;
 400         addr += 4;
 401     }
 402     /* define parameters */
 403     while ((sym = sym->next) != NULL) {
 404         u = sym->t;
 405         sym_push(sym->v & ~SYM_FIELD, u,
 406                  VT_LOCAL | VT_LVAL, addr);
 407         size = type_size(u, &align);
 408         size = (size + 3) & ~3;
 409 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 410         /* structs are passed as pointer */
 411         if ((u & VT_BTYPE) == VT_STRUCT) {
 412             size = 4;
 413         }
 414 #endif
 415         addr += size;
 416     }
 417     func_ret_sub = 0;
 418     /* pascal type call ? */
 419     if (func_call == FUNC_STDCALL)
 420         func_ret_sub = addr - 8;
 421     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 422     func_sub_sp_offset = oad(0xec81, 0); /* sub $xxx, %esp */
 423     /* leave some room for bound checking code */
 424     if (do_bounds_check) {
 425         oad(0xb8, 0); /* lbound section pointer */
 426         oad(0xb8, 0); /* call to function */
 427         func_bound_offset = lbounds_section->data_offset;
 428     }
 429 }
 430
 431 /* generate function epilog */
 432 void gfunc_epilog(void)
 433 {
 434 #ifdef CONFIG_TCC_BCHECK
 435     if (do_bounds_check && func_bound_offset != lbounds_section->data_offset) {
 436         int saved_ind;
 437         int *bounds_ptr;
 438         Sym *sym, *sym_data;
 439         /* add end of table info */
 440         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 441         *bounds_ptr = 0;
 442         /* generate bound local allocation */
 443         saved_ind = ind;
 444         ind = func_sub_sp_offset + 4;
 445         sym_data = get_sym_ref(char_pointer_type, lbounds_section,
 446                                func_bound_offset, lbounds_section->data_offset);
 447         greloc(cur_text_section, sym_data,
 448                ind + 1, R_386_32);
 449         oad(0xb8, 0); /* mov %eax, xxx */
 450         sym = external_global_sym(TOK___bound_local_new, func_old_type, 0);
 451         greloc(cur_text_section, sym,
 452                ind + 1, R_386_PC32);
 453         oad(0xe8, -4);
 454         ind = saved_ind;
 455         /* generate bound check local freeing */
 456         o(0x5250); /* save returned value, if any */
 457         greloc(cur_text_section, sym_data,
 458                ind + 1, R_386_32);
 459         oad(0xb8, 0); /* mov %eax, xxx */
 460         sym = external_global_sym(TOK___bound_local_delete, func_old_type, 0);
 461         greloc(cur_text_section, sym,
 462                ind + 1, R_386_PC32);
 463         oad(0xe8, -4);
 464         o(0x585a); /* restore returned value, if any */
 465     }
 466 #endif
 467     o(0xc9); /* leave */
 468     if (func_ret_sub == 0) {
 469         o(0xc3); /* ret */
 470     } else {
 471         o(0xc2); /* ret n */
 472         g(func_ret_sub);
 473         g(func_ret_sub >> 8);
 474     }
 475     /* align local size to word & save local variables */
 476     *(int *)(cur_text_section->data + func_sub_sp_offset) = (-loc + 3) & -4;
 477 }
 478
 479 /* generate a jump to a label */
 480 int gjmp(int t)
 481 {
 482     return psym(0xe9, t);
 483 }
 484
 485 /* generate a jump to a fixed address */
 486 void gjmp_addr(int a)
 487 {
 488     int r;
 489     r = a - ind - 2;
 490     if (r == (char)r) {
 491         g(0xeb);
 492         g(r);
 493     } else {
 494         oad(0xe9, a - ind - 5);
 495     }
 496 }
 497
 498 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 499 int gtst(int inv, int t)
 500 {
 501     int v, *p;
 502     v = vtop->r & VT_VALMASK;
 503     if (v == VT_CMP) {
 504         /* fast case : can jump directly since flags are set */
 505         g(0x0f);
 506         t = psym((vtop->c.i - 16) ^ inv, t);
 507     } else if (v == VT_JMP || v == VT_JMPI) {
 508         /* && or || optimization */
 509         if ((v & 1) == inv) {
 510             /* insert vtop->c jump list in t */
 511             p = &vtop->c.i;
 512             while (*p != 0)
 513                 p = (int *)(cur_text_section->data + *p);
 514             *p = t;
 515             t = vtop->c.i;
 516         } else {
 517             t = gjmp(t);
 518             gsym(vtop->c.i);
 519         }
 520     } else {
 521         if (is_float(vtop->t)) {
 522             vpushi(0);
 523             gen_op(TOK_NE);
 524         }
 525         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 526             /* constant jmp optimization */
 527             if ((vtop->c.i != 0) != inv)
 528                 t = gjmp(t);
 529         } else {
 530             v = gv(RC_INT);
 531             o(0x85);
 532             o(0xc0 + v * 9);
 533             g(0x0f);
 534             t = psym(0x85 ^ inv, t);
 535         }
 536     }
 537     vtop--;
 538     return t;
 539 }
 540
 541 /* generate an integer binary operation */
 542 void gen_opi(int op)
 543 {
 544     int r, fr, opc, c;
 545
 546     switch(op) {
 547     case '+':
 548     case TOK_ADDC1: /* add with carry generation */
 549         opc = 0;
 550     gen_op8:
 551         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 552             /* constant case */
 553             vswap();
 554             r = gv(RC_INT);
 555             vswap();
 556             c = vtop->c.i;
 557             if (c == (char)c) {
 558                 /* XXX: generate inc and dec for smaller code ? */
 559                 o(0x83);
 560                 o(0xc0 | (opc << 3) | r);
 561                 g(c);
 562             } else {
 563                 o(0x81);
 564                 oad(0xc0 | (opc << 3) | r, c);
 565             }
 566         } else {
 567             gv2(RC_INT, RC_INT);
 568             r = vtop[-1].r;
 569             fr = vtop[0].r;
 570             o((opc << 3) | 0x01);
 571             o(0xc0 + r + fr * 8);
 572         }
 573         vtop--;
 574         if (op >= TOK_ULT && op <= TOK_GT) {
 575             vtop--;
 576             vset(VT_INT, VT_CMP, op);
 577         }
 578         break;
 579     case '-':
 580     case TOK_SUBC1: /* sub with carry generation */
 581         opc = 5;
 582         goto gen_op8;
 583     case TOK_ADDC2: /* add with carry use */
 584         opc = 2;
 585         goto gen_op8;
 586     case TOK_SUBC2: /* sub with carry use */
 587         opc = 3;
 588         goto gen_op8;
 589     case '&':
 590         opc = 4;
 591         goto gen_op8;
 592     case '^':
 593         opc = 6;
 594         goto gen_op8;
 595     case '|':
 596         opc = 1;
 597         goto gen_op8;
 598     case '*':
 599         gv2(RC_INT, RC_INT);
 600         r = vtop[-1].r;
 601         fr = vtop[0].r;
 602         vtop--;
 603         o(0xaf0f); /* imul fr, r */
 604         o(0xc0 + fr + r * 8);
 605         break;
 606     case TOK_SHL:
 607         opc = 4;
 608         goto gen_shift;
 609     case TOK_SHR:
 610         opc = 5;
 611         goto gen_shift;
 612     case TOK_SAR:
 613         opc = 7;
 614     gen_shift:
 615         opc = 0xc0 | (opc << 3);
 616         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 617             /* constant case */
 618             vswap();
 619             r = gv(RC_INT);
 620             vswap();
 621             c = vtop->c.i & 0x1f;
 622             o(0xc1); /* shl/shr/sar $xxx, r */
 623             o(opc | r);
 624             g(c);
 625         } else {
 626             /* we generate the shift in ecx */
 627             gv2(RC_INT, RC_ECX);
 628             r = vtop[-1].r;
 629             o(0xd3); /* shl/shr/sar %cl, r */
 630             o(opc | r);
 631         }
 632         vtop--;
 633         break;
 634     case '/':
 635     case TOK_UDIV:
 636     case TOK_PDIV:
 637     case '%':
 638     case TOK_UMOD:
 639     case TOK_UMULL:
 640         /* first operand must be in eax */
 641         /* XXX: need better constraint for second operand */
 642         gv2(RC_EAX, RC_ECX);
 643         r = vtop[-1].r;
 644         fr = vtop[0].r;
 645         vtop--;
 646         save_reg(REG_EDX);
 647         if (op == TOK_UMULL) {
 648             o(0xf7); /* mul fr */
 649             o(0xe0 + fr);
 650             vtop->r2 = REG_EDX;
 651             r = REG_EAX;
 652         } else {
 653             if (op == TOK_UDIV || op == TOK_UMOD) {
 654                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 655                 o(0xf0 + fr);
 656             } else {
 657                 o(0xf799); /* cltd, idiv fr, %eax */
 658                 o(0xf8 + fr);
 659             }
 660             if (op == '%' || op == TOK_UMOD)
 661                 r = REG_EDX;
 662             else
 663                 r = REG_EAX;
 664         }
 665         vtop->r = r;
 666         break;
 667     default:
 668         opc = 7;
 669         goto gen_op8;
 670     }
 671 }
 672
 673 /* generate a floating point operation 'v = t1 op t2' instruction. The
 674    two operands are guaranted to have the same floating point type */
 675 /* XXX: need to use ST1 too */
 676 void gen_opf(int op)
 677 {
 678     int a, ft, fc, swapped, r;
 679
 680     /* convert constants to memory references */
 681     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 682         vswap();
 683         gv(RC_FLOAT);
 684         vswap();
 685     }
 686     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 687         gv(RC_FLOAT);
 688
 689     /* must put at least one value in the floating point register */
 690     if ((vtop[-1].r & VT_LVAL) &&
 691         (vtop[0].r & VT_LVAL)) {
 692         vswap();
 693         gv(RC_FLOAT);
 694         vswap();
 695     }
 696     swapped = 0;
 697     /* swap the stack if needed so that t1 is the register and t2 is
 698        the memory reference */
 699     if (vtop[-1].r & VT_LVAL) {
 700         vswap();
 701         swapped = 1;
 702     }
 703     if (op >= TOK_ULT && op <= TOK_GT) {
 704         /* load on stack second operand */
 705         load(REG_ST0, vtop);
 706         save_reg(REG_EAX); /* eax is used by FP comparison code */
 707         if (op == TOK_GE || op == TOK_GT)
 708             swapped = !swapped;
 709         else if (op == TOK_EQ || op == TOK_NE)
 710             swapped = 0;
 711         if (swapped)
 712             o(0xc9d9); /* fxch %st(1) */
 713         o(0xe9da); /* fucompp */
 714         o(0xe0df); /* fnstsw %ax */
 715         if (op == TOK_EQ) {
 716             o(0x45e480); /* and $0x45, %ah */
 717             o(0x40fC80); /* cmp $0x40, %ah */
 718         } else if (op == TOK_NE) {
 719             o(0x45e480); /* and $0x45, %ah */
 720             o(0x40f480); /* xor $0x40, %ah */
 721             op = TOK_NE;
 722         } else if (op == TOK_GE || op == TOK_LE) {
 723             o(0x05c4f6); /* test $0x05, %ah */
 724             op = TOK_EQ;
 725         } else {
 726             o(0x45c4f6); /* test $0x45, %ah */
 727             op = TOK_EQ;
 728         }
 729         vtop--;
 730         vtop->r = VT_CMP;
 731         vtop->c.i = op;
 732     } else {
 733         /* no memory reference possible for long double operations */
 734         if ((vtop->t & VT_BTYPE) == VT_LDOUBLE) {
 735             load(REG_ST0, vtop);
 736             swapped = !swapped;
 737         }
 738
 739         switch(op) {
 740         default:
 741         case '+':
 742             a = 0;
 743             break;
 744         case '-':
 745             a = 4;
 746             if (swapped)
 747                 a++;
 748             break;
 749         case '*':
 750             a = 1;
 751             break;
 752         case '/':
 753             a = 6;
 754             if (swapped)
 755                 a++;
 756             break;
 757         }
 758         ft = vtop->t;
 759         fc = vtop->c.ul;
 760         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 761             o(0xde); /* fxxxp %st, %st(1) */
 762             o(0xc1 + (a << 3));
 763         } else {
 764             /* if saved lvalue, then we must reload it */
 765             r = vtop->r;
 766             if ((r & VT_VALMASK) == VT_LLOCAL) {
 767                 SValue v1;
 768                 r = get_reg(RC_INT);
 769                 v1.t = VT_INT;
 770                 v1.r = VT_LOCAL | VT_LVAL;
 771                 v1.c.ul = fc;
 772                 load(r, &v1);
 773                 fc = 0;
 774             }
 775
 776             if ((ft & VT_BTYPE) == VT_DOUBLE)
 777                 o(0xdc);
 778             else
 779                 o(0xd8);
 780             gen_modrm(a, r, vtop->sym, fc);
 781         }
 782         vtop--;
 783     }
 784 }
 785
 786 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 787    and 'long long' cases. */
 788 void gen_cvt_itof(int t)
 789 {
 790     save_reg(REG_ST0);
 791     gv(RC_INT);
 792     if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 793         /* signed long long to float/double/long double (unsigned case
 794            is handled generically) */
 795         o(0x50 + vtop->r2); /* push r2 */
 796         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 797         o(0x242cdf); /* fildll (%esp) */
 798         o(0x08c483); /* add $8, %esp */
 799     } else if ((vtop->t & (VT_BTYPE | VT_UNSIGNED)) ==
 800                (VT_INT | VT_UNSIGNED)) {
 801         /* unsigned int to float/double/long double */
 802         o(0x6a); /* push $0 */
 803         g(0x00);
 804         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 805         o(0x242cdf); /* fildll (%esp) */
 806         o(0x08c483); /* add $8, %esp */
 807     } else {
 808         /* int to float/double/long double */
 809         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 810         o(0x2404db); /* fildl (%esp) */
 811         o(0x04c483); /* add $4, %esp */
 812     }
 813     vtop->r = REG_ST0;
 814 }
 815
 816 /* convert fp to int 't' type */
 817 /* XXX: handle long long case */
 818 void gen_cvt_ftoi(int t)
 819 {
 820     int r, r2, size;
 821     Sym *sym;
 822
 823     gv(RC_FLOAT);
 824     if (t != VT_INT)
 825         size = 8;
 826     else
 827         size = 4;
 828
 829     o(0x2dd9); /* ldcw xxx */
 830     sym = external_global_sym(TOK___tcc_int_fpu_control,
 831                               VT_SHORT | VT_UNSIGNED, VT_LVAL);
 832     greloc(cur_text_section, sym,
 833            ind, R_386_32);
 834     gen_le32(0);
 835
 836     oad(0xec81, size); /* sub $xxx, %esp */
 837     if (size == 4)
 838         o(0x1cdb); /* fistpl */
 839     else
 840         o(0x3cdf); /* fistpll */
 841     o(0x24);
 842     o(0x2dd9); /* ldcw xxx */
 843     sym = external_global_sym(TOK___tcc_fpu_control,
 844                               VT_SHORT | VT_UNSIGNED, VT_LVAL);
 845     greloc(cur_text_section, sym,
 846            ind, R_386_32);
 847     gen_le32(0);
 848
 849     r = get_reg(RC_INT);
 850     o(0x58 + r); /* pop r */
 851     if (size == 8) {
 852         if (t == VT_LLONG) {
 853             vtop->r = r; /* mark reg as used */
 854             r2 = get_reg(RC_INT);
 855             o(0x58 + r2); /* pop r2 */
 856             vtop->r2 = r2;
 857         } else {
 858             o(0x04c483); /* add $4, %esp */
 859         }
 860     }
 861     vtop->r = r;
 862 }
 863
 864 /* convert from one floating point type to another */
 865 void gen_cvt_ftof(int t)
 866 {
 867     /* all we have to do on i386 is to put the float in a register */
 868     gv(RC_FLOAT);
 869 }
 870
 871 /* computed goto support */
 872 void ggoto(void)
 873 {
 874     gcall_or_jmp(1);
 875     vtop--;
 876 }
 877
 878 /* bound check support functions */
 879 #ifdef CONFIG_TCC_BCHECK
 880
 881 /* generate a bounded pointer addition */
 882 void gen_bounded_ptr_add(void)
 883 {
 884     Sym *sym;
 885
 886     /* prepare fast i386 function call (args in eax and edx) */
 887     gv2(RC_EAX, RC_EDX);
 888     /* save all temporary registers */
 889     vtop -= 2;
 890     save_regs(0);
 891     /* do a fast function call */
 892     sym = external_global_sym(TOK___bound_ptr_add, func_old_type, 0);
 893     greloc(cur_text_section, sym,
 894            ind + 1, R_386_PC32);
 895     oad(0xe8, -4);
 896     /* returned pointer is in eax */
 897     vtop++;
 898     vtop->r = REG_EAX | VT_BOUNDED;
 899     /* address of bounding function call point */
 900     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 901 }
 902
 903 /* patch pointer addition in vtop so that pointer dereferencing is
 904    also tested */
 905 void gen_bounded_ptr_deref(void)
 906 {
 907     int func;
 908     int size, align;
 909     Elf32_Rel *rel;
 910     Sym *sym;
 911
 912     size = 0;
 913     /* XXX: put that code in generic part of tcc */
 914     if (!is_float(vtop->t)) {
 915         if (vtop->r & VT_LVAL_BYTE)
 916             size = 1;
 917         else if (vtop->r & VT_LVAL_SHORT)
 918             size = 2;
 919     }
 920     if (!size)
 921         size = type_size(vtop->t, &align);
 922     switch(size) {
 923     case  1: func = TOK___bound_ptr_indir1; break;
 924     case  2: func = TOK___bound_ptr_indir2; break;
 925     case  4: func = TOK___bound_ptr_indir4; break;
 926     case  8: func = TOK___bound_ptr_indir8; break;
 927     case 12: func = TOK___bound_ptr_indir12; break;
 928     case 16: func = TOK___bound_ptr_indir16; break;
 929     default:
 930         error("unhandled size when derefencing bounded pointer");
 931         func = 0;
 932         break;
 933     }
 934
 935     /* patch relocation */
 936     /* XXX: find a better solution ? */
 937     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
 938     sym = external_global_sym(func, func_old_type, 0);
 939     if (!sym->c)
 940         put_extern_sym(sym, NULL, 0, 0);
 941     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
 942 }
 943 #endif
 944
 945 /* end of X86 code generator */
 946 /*************************************************************/
 947