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 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 void gen_addr32(int r, int c)
 151 {
 152     if (!(r & VT_SYM)) {
 153         gen_le32(c);
 154     } else {
 155         greloc(cur_text_section,
 156                (Sym *)c, ind, R_386_32);
 157         gen_le32(0);
 158     }
 159 }
 160
 161 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 162    opcode bits */
 163 void gen_modrm(int op_reg, int r, int c)
 164 {
 165     op_reg = op_reg << 3;
 166     if ((r & VT_VALMASK) == VT_CONST) {
 167         /* constant memory reference */
 168         o(0x05 | op_reg);
 169         gen_addr32(r, c);
 170     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 171         /* currently, we use only ebp as base */
 172         if (c == (char)c) {
 173             /* short reference */
 174             o(0x45 | op_reg);
 175             g(c);
 176         } else {
 177             oad(0x85 | op_reg, c);
 178         }
 179     } else {
 180         g(0x00 | op_reg | (r & VT_VALMASK));
 181     }
 182 }
 183
 184
 185 /* load 'r' from value 'sv' */
 186 void load(int r, SValue *sv)
 187 {
 188     int v, t, ft, fc, fr;
 189     SValue v1;
 190
 191     fr = sv->r;
 192     ft = sv->t;
 193     fc = sv->c.ul;
 194
 195     v = fr & VT_VALMASK;
 196     if (fr & VT_LVAL) {
 197         if (v == VT_LLOCAL) {
 198             v1.t = VT_INT;
 199             v1.r = VT_LOCAL | VT_LVAL;
 200             v1.c.ul = fc;
 201             load(r, &v1);
 202             fr = r;
 203         }
 204         if ((ft & VT_BTYPE) == VT_FLOAT) {
 205             o(0xd9); /* flds */
 206             r = 0;
 207         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 208             o(0xdd); /* fldl */
 209             r = 0;
 210         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 211             o(0xdb); /* fldt */
 212             r = 5;
 213         } else if ((ft & VT_TYPE) == VT_BYTE) {
 214             o(0xbe0f);   /* movsbl */
 215         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 216             o(0xb60f);   /* movzbl */
 217         } else if ((ft & VT_TYPE) == VT_SHORT) {
 218             o(0xbf0f);   /* movswl */
 219         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 220             o(0xb70f);   /* movzwl */
 221         } else {
 222             o(0x8b);     /* movl */
 223         }
 224         gen_modrm(r, fr, fc);
 225     } else {
 226         if (v == VT_CONST) {
 227             o(0xb8 + r); /* mov $xx, r */
 228             gen_addr32(fr, fc);
 229         } else if (v == VT_LOCAL) {
 230             o(0x8d); /* lea xxx(%ebp), r */
 231             gen_modrm(r, VT_LOCAL, fc);
 232         } else if (v == VT_CMP) {
 233             oad(0xb8 + r, 0); /* mov $0, r */
 234             o(0x0f); /* setxx %br */
 235             o(fc);
 236             o(0xc0 + r);
 237         } else if (v == VT_JMP || v == VT_JMPI) {
 238             t = v & 1;
 239             oad(0xb8 + r, t); /* mov $1, r */
 240             oad(0xe9, 5); /* jmp after */
 241             gsym(fc);
 242             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 243         } else if (v != r) {
 244             o(0x89);
 245             o(0xc0 + r + v * 8); /* mov v, r */
 246         }
 247     }
 248 }
 249
 250 /* store register 'r' in lvalue 'v' */
 251 void store(int r, SValue *v)
 252 {
 253     int fr, bt, ft, fc;
 254
 255     ft = v->t;
 256     fc = v->c.ul;
 257     fr = v->r & VT_VALMASK;
 258     bt = ft & VT_BTYPE;
 259     /* XXX: incorrect if float reg to reg */
 260     if (bt == VT_FLOAT) {
 261         o(0xd9); /* fsts */
 262         r = 2;
 263     } else if (bt == VT_DOUBLE) {
 264         o(0xdd); /* fstpl */
 265         r = 2;
 266     } else if (bt == VT_LDOUBLE) {
 267         o(0xc0d9); /* fld %st(0) */
 268         o(0xdb); /* fstpt */
 269         r = 7;
 270     } else {
 271         if (bt == VT_SHORT)
 272             o(0x66);
 273         if (bt == VT_BYTE)
 274             o(0x88);
 275         else
 276             o(0x89);
 277     }
 278     if (fr == VT_CONST ||
 279         fr == VT_LOCAL ||
 280         (v->r & VT_LVAL)) {
 281         gen_modrm(r, v->r, fc);
 282     } else if (fr != r) {
 283         o(0xc0 + fr + r * 8); /* mov r, fr */
 284     }
 285 }
 286
 287 /* start function call and return function call context */
 288 void gfunc_start(GFuncContext *c, int func_call)
 289 {
 290     c->args_size = 0;
 291     c->func_call = func_call;
 292 }
 293
 294 /* push function parameter which is in (vtop->t, vtop->c). Stack entry
 295    is then popped. */
 296 void gfunc_param(GFuncContext *c)
 297 {
 298     int size, align, r;
 299
 300     if ((vtop->t & VT_BTYPE) == VT_STRUCT) {
 301         size = type_size(vtop->t, &align);
 302         /* align to stack align size */
 303         size = (size + 3) & ~3;
 304         /* allocate the necessary size on stack */
 305         oad(0xec81, size); /* sub $xxx, %esp */
 306         /* generate structure store */
 307         r = get_reg(RC_INT);
 308         o(0x89); /* mov %esp, r */
 309         o(0xe0 + r);
 310         vset(vtop->t, r | VT_LVAL, 0);
 311         vswap();
 312         vstore();
 313         c->args_size += size;
 314     } else if (is_float(vtop->t)) {
 315         gv(RC_FLOAT); /* only one float register */
 316         if ((vtop->t & VT_BTYPE) == VT_FLOAT)
 317             size = 4;
 318         else if ((vtop->t & VT_BTYPE) == VT_DOUBLE)
 319             size = 8;
 320         else
 321             size = 12;
 322         oad(0xec81, size); /* sub $xxx, %esp */
 323         if (size == 12)
 324             o(0x7cdb);
 325         else
 326             o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 327         g(0x24);
 328         g(0x00);
 329         c->args_size += size;
 330     } else {
 331         /* simple type (currently always same size) */
 332         /* XXX: implicit cast ? */
 333         r = gv(RC_INT);
 334         if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 335             size = 8;
 336             o(0x50 + vtop->r2); /* push r */
 337         } else {
 338             size = 4;
 339         }
 340         o(0x50 + r); /* push r */
 341         c->args_size += size;
 342     }
 343     vtop--;
 344 }
 345
 346 static void gadd_sp(int val)
 347 {
 348     if (val == (char)val) {
 349         o(0xc483);
 350         g(val);
 351     } else {
 352         oad(0xc481, val); /* add $xxx, %esp */
 353     }
 354 }
 355
 356 /* generate function call with address in (vtop->t, vtop->c) and free function
 357    context. Stack entry is popped */
 358 void gfunc_call(GFuncContext *c)
 359 {
 360     int r;
 361     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 362         /* constant case */
 363         if (vtop->r & VT_SYM) {
 364             /* relocation case */
 365             greloc(cur_text_section, vtop->c.sym,
 366                    ind + 1, R_386_PC32);
 367             oad(0xe8, -4);
 368         } else {
 369             oad(0xe8, vtop->c.ul - ind - 5);
 370         }
 371     } else {
 372         /* otherwise, indirect call */
 373         r = gv(RC_INT);
 374         o(0xff); /* call *r */
 375         o(0xd0 + r);
 376     }
 377     if (c->args_size && c->func_call == FUNC_CDECL)
 378         gadd_sp(c->args_size);
 379     vtop--;
 380 }
 381
 382 /* generate function prolog of type 't' */
 383 void gfunc_prolog(int t)
 384 {
 385     int addr, align, size, u, func_call;
 386     Sym *sym;
 387
 388     sym = sym_find((unsigned)t >> VT_STRUCT_SHIFT);
 389     func_call = sym->r;
 390     addr = 8;
 391     /* if the function returns a structure, then add an
 392        implicit pointer parameter */
 393     func_vt = sym->t;
 394     if ((func_vt & VT_BTYPE) == VT_STRUCT) {
 395         func_vc = addr;
 396         addr += 4;
 397     }
 398     /* define parameters */
 399     while ((sym = sym->next) != NULL) {
 400         u = sym->t;
 401         sym_push(sym->v & ~SYM_FIELD, u,
 402                  VT_LOCAL | VT_LVAL, addr);
 403         size = type_size(u, &align);
 404         size = (size + 3) & ~3;
 405 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 406         /* structs are passed as pointer */
 407         if ((u & VT_BTYPE) == VT_STRUCT) {
 408             size = 4;
 409         }
 410 #endif
 411         addr += size;
 412     }
 413     func_ret_sub = 0;
 414     /* pascal type call ? */
 415     if (func_call == FUNC_STDCALL)
 416         func_ret_sub = addr - 8;
 417     o(0xe58955); /* push   %ebp, mov    %esp, %ebp */
 418     func_sub_sp_offset = oad(0xec81, 0); /* sub $xxx, %esp */
 419     /* leave some room for bound checking code */
 420     if (do_bounds_check) {
 421         oad(0xb8, 0); /* lbound section pointer */
 422         oad(0xb8, 0); /* call to function */
 423         func_bound_offset = lbounds_section->data_offset;
 424     }
 425 }
 426
 427 /* generate function epilog */
 428 void gfunc_epilog(void)
 429 {
 430 #ifdef CONFIG_TCC_BCHECK
 431     if (do_bounds_check && func_bound_offset != lbounds_section->data_offset) {
 432         int saved_ind;
 433         int *bounds_ptr;
 434         Sym *sym, *sym_data;
 435         /* add end of table info */
 436         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 437         *bounds_ptr = 0;
 438         /* generate bound local allocation */
 439         saved_ind = ind;
 440         ind = func_sub_sp_offset + 4;
 441         sym_data = get_sym_ref(char_pointer_type, lbounds_section,
 442                                func_bound_offset, lbounds_section->data_offset);
 443         greloc(cur_text_section, sym_data,
 444                ind + 1, R_386_32);
 445         oad(0xb8, 0); /* mov %eax, xxx */
 446         sym = external_sym(TOK___bound_local_new, func_old_type, 0);
 447         greloc(cur_text_section, sym,
 448                ind + 1, R_386_PC32);
 449         oad(0xe8, -4);
 450         ind = saved_ind;
 451         /* generate bound check local freeing */
 452         o(0x5250); /* save returned value, if any */
 453         greloc(cur_text_section, sym_data,
 454                ind + 1, R_386_32);
 455         oad(0xb8, 0); /* mov %eax, xxx */
 456         sym = external_sym(TOK___bound_local_delete, func_old_type, 0);
 457         greloc(cur_text_section, sym,
 458                ind + 1, R_386_PC32);
 459         oad(0xe8, -4);
 460         o(0x585a); /* restore returned value, if any */
 461     }
 462 #endif
 463     o(0xc9); /* leave */
 464     if (func_ret_sub == 0) {
 465         o(0xc3); /* ret */
 466     } else {
 467         o(0xc2); /* ret n */
 468         g(func_ret_sub);
 469         g(func_ret_sub >> 8);
 470     }
 471     /* align local size to word & save local variables */
 472     *(int *)(cur_text_section->data + func_sub_sp_offset) = (-loc + 3) & -4;
 473 }
 474
 475 /* generate a jump to a label */
 476 int gjmp(int t)
 477 {
 478     return psym(0xe9, t);
 479 }
 480
 481 /* generate a jump to a fixed address */
 482 void gjmp_addr(int a)
 483 {
 484     oad(0xe9, a - ind - 5);
 485 }
 486
 487 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 488 int gtst(int inv, int t)
 489 {
 490     int v, *p;
 491     v = vtop->r & VT_VALMASK;
 492     if (v == VT_CMP) {
 493         /* fast case : can jump directly since flags are set */
 494         g(0x0f);
 495         t = psym((vtop->c.i - 16) ^ inv, t);
 496     } else if (v == VT_JMP || v == VT_JMPI) {
 497         /* && or || optimization */
 498         if ((v & 1) == inv) {
 499             /* insert vtop->c jump list in t */
 500             p = &vtop->c.i;
 501             while (*p != 0)
 502                 p = (int *)(cur_text_section->data + *p);
 503             *p = t;
 504             t = vtop->c.i;
 505         } else {
 506             t = gjmp(t);
 507             gsym(vtop->c.i);
 508         }
 509     } else {
 510         if (is_float(vtop->t)) {
 511             vpushi(0);
 512             gen_op(TOK_NE);
 513         }
 514         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 515             /* constant jmp optimization */
 516             if ((vtop->c.i != 0) != inv)
 517                 t = gjmp(t);
 518         } else {
 519             v = gv(RC_INT);
 520             o(0x85);
 521             o(0xc0 + v * 9);
 522             g(0x0f);
 523             t = psym(0x85 ^ inv, t);
 524         }
 525     }
 526     vtop--;
 527     return t;
 528 }
 529
 530 /* generate an integer binary operation */
 531 void gen_opi(int op)
 532 {
 533     int r, fr, opc, c;
 534
 535     switch(op) {
 536     case '+':
 537     case TOK_ADDC1: /* add with carry generation */
 538         opc = 0;
 539     gen_op8:
 540         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 541             /* constant case */
 542             vswap();
 543             r = gv(RC_INT);
 544             vswap();
 545             c = vtop->c.i;
 546             if (c == (char)c) {
 547                 /* XXX: generate inc and dec for smaller code ? */
 548                 o(0x83);
 549                 o(0xc0 | (opc << 3) | r);
 550                 g(c);
 551             } else {
 552                 o(0x81);
 553                 oad(0xc0 | (opc << 3) | r, c);
 554             }
 555         } else {
 556             gv2(RC_INT, RC_INT);
 557             r = vtop[-1].r;
 558             fr = vtop[0].r;
 559             o((opc << 3) | 0x01);
 560             o(0xc0 + r + fr * 8);
 561         }
 562         vtop--;
 563         if (op >= TOK_ULT && op <= TOK_GT) {
 564             vtop--;
 565             vset(VT_INT, VT_CMP, op);
 566         }
 567         break;
 568     case '-':
 569     case TOK_SUBC1: /* sub with carry generation */
 570         opc = 5;
 571         goto gen_op8;
 572     case TOK_ADDC2: /* add with carry use */
 573         opc = 2;
 574         goto gen_op8;
 575     case TOK_SUBC2: /* sub with carry use */
 576         opc = 3;
 577         goto gen_op8;
 578     case '&':
 579         opc = 4;
 580         goto gen_op8;
 581     case '^':
 582         opc = 6;
 583         goto gen_op8;
 584     case '|':
 585         opc = 1;
 586         goto gen_op8;
 587     case '*':
 588         gv2(RC_INT, RC_INT);
 589         r = vtop[-1].r;
 590         fr = vtop[0].r;
 591         vtop--;
 592         o(0xaf0f); /* imul fr, r */
 593         o(0xc0 + fr + r * 8);
 594         break;
 595     case TOK_SHL:
 596         opc = 4;
 597         goto gen_shift;
 598     case TOK_SHR:
 599         opc = 5;
 600         goto gen_shift;
 601     case TOK_SAR:
 602         opc = 7;
 603     gen_shift:
 604         opc = 0xc0 | (opc << 3);
 605         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 606             /* constant case */
 607             vswap();
 608             r = gv(RC_INT);
 609             vswap();
 610             c = vtop->c.i & 0x1f;
 611             o(0xc1); /* shl/shr/sar $xxx, r */
 612             o(opc | r);
 613             g(c);
 614         } else {
 615             /* we generate the shift in ecx */
 616             gv2(RC_INT, RC_ECX);
 617             r = vtop[-1].r;
 618             o(0xd3); /* shl/shr/sar %cl, r */
 619             o(opc | r);
 620         }
 621         vtop--;
 622         break;
 623     case '/':
 624     case TOK_UDIV:
 625     case TOK_PDIV:
 626     case '%':
 627     case TOK_UMOD:
 628     case TOK_UMULL:
 629         /* first operand must be in eax */
 630         /* XXX: need better constraint for second operand */
 631         gv2(RC_EAX, RC_ECX);
 632         r = vtop[-1].r;
 633         fr = vtop[0].r;
 634         vtop--;
 635         save_reg(REG_EDX);
 636         if (op == TOK_UMULL) {
 637             o(0xf7); /* mul fr */
 638             o(0xe0 + fr);
 639             vtop->r2 = REG_EDX;
 640             r = REG_EAX;
 641         } else {
 642             if (op == TOK_UDIV || op == TOK_UMOD) {
 643                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 644                 o(0xf0 + fr);
 645             } else {
 646                 o(0xf799); /* cltd, idiv fr, %eax */
 647                 o(0xf8 + fr);
 648             }
 649             if (op == '%' || op == TOK_UMOD)
 650                 r = REG_EDX;
 651             else
 652                 r = REG_EAX;
 653         }
 654         vtop->r = r;
 655         break;
 656     default:
 657         opc = 7;
 658         goto gen_op8;
 659     }
 660 }
 661
 662 /* generate a floating point operation 'v = t1 op t2' instruction. The
 663    two operands are guaranted to have the same floating point type */
 664 /* XXX: need to use ST1 too */
 665 void gen_opf(int op)
 666 {
 667     int a, ft, fc, swapped, r;
 668
 669     /* convert constants to memory references */
 670     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 671         vswap();
 672         gv(RC_FLOAT);
 673         vswap();
 674     }
 675     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 676         gv(RC_FLOAT);
 677
 678     /* must put at least one value in the floating point register */
 679     if ((vtop[-1].r & VT_LVAL) &&
 680         (vtop[0].r & VT_LVAL)) {
 681         vswap();
 682         gv(RC_FLOAT);
 683         vswap();
 684     }
 685     swapped = 0;
 686     /* swap the stack if needed so that t1 is the register and t2 is
 687        the memory reference */
 688     if (vtop[-1].r & VT_LVAL) {
 689         vswap();
 690         swapped = 1;
 691     }
 692     if (op >= TOK_ULT && op <= TOK_GT) {
 693         /* load on stack second operand */
 694         load(REG_ST0, vtop);
 695         save_reg(REG_EAX); /* eax is used by FP comparison code */
 696         if (op == TOK_GE || op == TOK_GT)
 697             swapped = !swapped;
 698         else if (op == TOK_EQ || op == TOK_NE)
 699             swapped = 0;
 700         if (swapped)
 701             o(0xc9d9); /* fxch %st(1) */
 702         o(0xe9da); /* fucompp */
 703         o(0xe0df); /* fnstsw %ax */
 704         if (op == TOK_EQ) {
 705             o(0x45e480); /* and $0x45, %ah */
 706             o(0x40fC80); /* cmp $0x40, %ah */
 707         } else if (op == TOK_NE) {
 708             o(0x45e480); /* and $0x45, %ah */
 709             o(0x40f480); /* xor $0x40, %ah */
 710             op = TOK_NE;
 711         } else if (op == TOK_GE || op == TOK_LE) {
 712             o(0x05c4f6); /* test $0x05, %ah */
 713             op = TOK_EQ;
 714         } else {
 715             o(0x45c4f6); /* test $0x45, %ah */
 716             op = TOK_EQ;
 717         }
 718         vtop--;
 719         vtop->r = VT_CMP;
 720         vtop->c.i = op;
 721     } else {
 722         /* no memory reference possible for long double operations */
 723         if ((vtop->t & VT_BTYPE) == VT_LDOUBLE) {
 724             load(REG_ST0, vtop);
 725             swapped = !swapped;
 726         }
 727
 728         switch(op) {
 729         default:
 730         case '+':
 731             a = 0;
 732             break;
 733         case '-':
 734             a = 4;
 735             if (swapped)
 736                 a++;
 737             break;
 738         case '*':
 739             a = 1;
 740             break;
 741         case '/':
 742             a = 6;
 743             if (swapped)
 744                 a++;
 745             break;
 746         }
 747         ft = vtop->t;
 748         fc = vtop->c.ul;
 749         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 750             o(0xde); /* fxxxp %st, %st(1) */
 751             o(0xc1 + (a << 3));
 752         } else {
 753             /* if saved lvalue, then we must reload it */
 754             r = vtop->r;
 755             if ((r & VT_VALMASK) == VT_LLOCAL) {
 756                 SValue v1;
 757                 r = get_reg(RC_INT);
 758                 v1.t = VT_INT;
 759                 v1.r = VT_LOCAL | VT_LVAL;
 760                 v1.c.ul = fc;
 761                 load(r, &v1);
 762                 fc = 0;
 763             }
 764
 765             if ((ft & VT_BTYPE) == VT_DOUBLE)
 766                 o(0xdc);
 767             else
 768                 o(0xd8);
 769             gen_modrm(a, r, fc);
 770         }
 771         vtop--;
 772     }
 773 }
 774
 775 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 776    and 'long long' cases. */
 777 void gen_cvt_itof(int t)
 778 {
 779     save_reg(REG_ST0);
 780     gv(RC_INT);
 781     if ((vtop->t & VT_BTYPE) == VT_LLONG) {
 782         /* signed long long to float/double/long double (unsigned case
 783            is handled generically) */
 784         o(0x50 + vtop->r2); /* push r2 */
 785         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 786         o(0x242cdf); /* fildll (%esp) */
 787         o(0x08c483); /* add $8, %esp */
 788     } else if ((vtop->t & (VT_BTYPE | VT_UNSIGNED)) ==
 789                (VT_INT | VT_UNSIGNED)) {
 790         /* unsigned int to float/double/long double */
 791         o(0x6a); /* push $0 */
 792         g(0x00);
 793         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 794         o(0x242cdf); /* fildll (%esp) */
 795         o(0x08c483); /* add $8, %esp */
 796     } else {
 797         /* int to float/double/long double */
 798         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 799         o(0x2404db); /* fildl (%esp) */
 800         o(0x04c483); /* add $4, %esp */
 801     }
 802     vtop->r = REG_ST0;
 803 }
 804
 805 /* convert fp to int 't' type */
 806 /* XXX: handle long long case */
 807 void gen_cvt_ftoi(int t)
 808 {
 809     int r, r2, size;
 810     Sym *sym;
 811
 812     gv(RC_FLOAT);
 813     if (t != VT_INT)
 814         size = 8;
 815     else
 816         size = 4;
 817
 818     o(0x2dd9); /* ldcw xxx */
 819     sym = external_sym(TOK___tcc_int_fpu_control,
 820                        VT_SHORT | VT_UNSIGNED, VT_LVAL);
 821     greloc(cur_text_section, sym,
 822            ind, R_386_32);
 823     gen_le32(0);
 824
 825     oad(0xec81, size); /* sub $xxx, %esp */
 826     if (size == 4)
 827         o(0x1cdb); /* fistpl */
 828     else
 829         o(0x3cdf); /* fistpll */
 830     o(0x24);
 831     o(0x2dd9); /* ldcw xxx */
 832     sym = external_sym(TOK___tcc_fpu_control,
 833                        VT_SHORT | VT_UNSIGNED, VT_LVAL);
 834     greloc(cur_text_section, sym,
 835            ind, R_386_32);
 836     gen_le32(0);
 837
 838     r = get_reg(RC_INT);
 839     o(0x58 + r); /* pop r */
 840     if (size == 8) {
 841         if (t == VT_LLONG) {
 842             vtop->r = r; /* mark reg as used */
 843             r2 = get_reg(RC_INT);
 844             o(0x58 + r2); /* pop r2 */
 845             vtop->r2 = r2;
 846         } else {
 847             o(0x04c483); /* add $4, %esp */
 848         }
 849     }
 850     vtop->r = r;
 851 }
 852
 853 /* convert from one floating point type to another */
 854 void gen_cvt_ftof(int t)
 855 {
 856     /* all we have to do on i386 is to put the float in a register */
 857     gv(RC_FLOAT);
 858 }
 859
 860 /* bound check support functions */
 861 #ifdef CONFIG_TCC_BCHECK
 862
 863 /* generate a bounded pointer addition */
 864 void gen_bounded_ptr_add(void)
 865 {
 866     Sym *sym;
 867
 868     /* prepare fast i386 function call (args in eax and edx) */
 869     gv2(RC_EAX, RC_EDX);
 870     /* save all temporary registers */
 871     vtop -= 2;
 872     save_regs(0);
 873     /* do a fast function call */
 874     sym = external_sym(TOK___bound_ptr_add, func_old_type, 0);
 875     greloc(cur_text_section, sym,
 876            ind + 1, R_386_PC32);
 877     oad(0xe8, -4);
 878     /* returned pointer is in eax */
 879     vtop++;
 880     vtop->r = REG_EAX | VT_BOUNDED;
 881     /* address of bounding function call point */
 882     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 883 }
 884
 885 /* patch pointer addition in vtop so that pointer dereferencing is
 886    also tested */
 887 void gen_bounded_ptr_deref(void)
 888 {
 889     int func;
 890     int size, align;
 891     Elf32_Rel *rel;
 892     Sym *sym;
 893
 894     size = 0;
 895     /* XXX: put that code in generic part of tcc */
 896     if (!is_float(vtop->t)) {
 897         if (vtop->r & VT_LVAL_BYTE)
 898             size = 1;
 899         else if (vtop->r & VT_LVAL_SHORT)
 900             size = 2;
 901     }
 902     if (!size)
 903         size = type_size(vtop->t, &align);
 904     switch(size) {
 905     case  1: func = TOK___bound_ptr_indir1; break;
 906     case  2: func = TOK___bound_ptr_indir2; break;
 907     case  4: func = TOK___bound_ptr_indir4; break;
 908     case  8: func = TOK___bound_ptr_indir8; break;
 909     case 12: func = TOK___bound_ptr_indir12; break;
 910     case 16: func = TOK___bound_ptr_indir16; break;
 911     default:
 912         error("unhandled size when derefencing bounded pointer");
 913         func = 0;
 914         break;
 915     }
 916
 917     /* patch relocation */
 918     /* XXX: find a better solution ? */
 919     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
 920     sym = external_sym(func, func_old_type, 0);
 921     if (!sym->c)
 922         put_extern_sym(sym, NULL, 0, 0);
 923     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
 924 }
 925 #endif
 926
 927 /* end of X86 code generator */
 928 /*************************************************************/
 929