i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001-2004 Fabrice Bellard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  19  */
  20
  21 /* number of available registers */
  22 #define NB_REGS             4
  23
  24 /* a register can belong to several classes. The classes must be
  25    sorted from more general to more precise (see gv2() code which does
  26    assumptions on it). */
  27 #define RC_INT     0x0001 /* generic integer register */
  28 #define RC_FLOAT   0x0002 /* generic float register */
  29 #define RC_EAX     0x0004
  30 #define RC_ST0     0x0008
  31 #define RC_ECX     0x0010
  32 #define RC_EDX     0x0020
  33 #define RC_IRET    RC_EAX /* function return: integer register */
  34 #define RC_LRET    RC_EDX /* function return: second integer register */
  35 #define RC_FRET    RC_ST0 /* function return: float register */
  36
  37 /* pretty names for the registers */
  38 enum {
  39     TREG_EAX = 0,
  40     TREG_ECX,
  41     TREG_EDX,
  42     TREG_ST0,
  43 };
  44
  45 int reg_classes[NB_REGS] = {
  46     /* eax */ RC_INT | RC_EAX,
  47     /* ecx */ RC_INT | RC_ECX,
  48     /* edx */ RC_INT | RC_EDX,
  49     /* st0 */ RC_FLOAT | RC_ST0,
  50 };
  51
  52 /* return registers for function */
  53 #define REG_IRET TREG_EAX /* single word int return register */
  54 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  55 #define REG_FRET TREG_ST0 /* float return register */
  56
  57 /* defined if function parameters must be evaluated in reverse order */
  58 #define INVERT_FUNC_PARAMS
  59
  60 /* defined if structures are passed as pointers. Otherwise structures
  61    are directly pushed on stack. */
  62 //#define FUNC_STRUCT_PARAM_AS_PTR
  63
  64 /* pointer size, in bytes */
  65 #define PTR_SIZE 4
  66
  67 /* long double size and alignment, in bytes */
  68 #define LDOUBLE_SIZE  12
  69 #define LDOUBLE_ALIGN 4
  70 /* maximum alignment (for aligned attribute support) */
  71 #define MAX_ALIGN     8
  72
  73 /******************************************************/
  74 /* ELF defines */
  75
  76 #define EM_TCC_TARGET EM_386
  77
  78 /* relocation type for 32 bit data relocation */
  79 #define R_DATA_32   R_386_32
  80 #define R_JMP_SLOT  R_386_JMP_SLOT
  81 #define R_COPY      R_386_COPY
  82
  83 #define ELF_START_ADDR 0x08048000
  84 #define ELF_PAGE_SIZE  0x1000
  85
  86 /******************************************************/
  87
  88 static unsigned long func_sub_sp_offset;
  89 static unsigned long func_bound_offset;
  90 static int func_ret_sub;
  91
  92 /* XXX: make it faster ? */
  93 void g(int c)
  94 {
  95     int ind1;
  96     ind1 = ind + 1;
  97     if (ind1 > cur_text_section->data_allocated)
  98         section_realloc(cur_text_section, ind1);
  99     cur_text_section->data[ind] = c;
 100     ind = ind1;
 101 }
 102
 103 void o(unsigned int c)
 104 {
 105     while (c) {
 106         g(c);
 107         c = c >> 8;
 108     }
 109 }
 110
 111 void gen_le32(int c)
 112 {
 113     g(c);
 114     g(c >> 8);
 115     g(c >> 16);
 116     g(c >> 24);
 117 }
 118
 119 /* output a symbol and patch all calls to it */
 120 void gsym_addr(int t, int a)
 121 {
 122     int n, *ptr;
 123     while (t) {
 124         ptr = (int *)(cur_text_section->data + t);
 125         n = *ptr; /* next value */
 126         *ptr = a - t - 4;
 127         t = n;
 128     }
 129 }
 130
 131 void gsym(int t)
 132 {
 133     gsym_addr(t, ind);
 134 }
 135
 136 /* psym is used to put an instruction with a data field which is a
 137    reference to a symbol. It is in fact the same as oad ! */
 138 #define psym oad
 139
 140 /* instruction + 4 bytes data. Return the address of the data */
 141 static int oad(int c, int s)
 142 {
 143     int ind1;
 144
 145     o(c);
 146     ind1 = ind + 4;
 147     if (ind1 > cur_text_section->data_allocated)
 148         section_realloc(cur_text_section, ind1);
 149     *(int *)(cur_text_section->data + ind) = s;
 150     s = ind;
 151     ind = ind1;
 152     return s;
 153 }
 154
 155 /* output constant with relocation if 'r & VT_SYM' is true */
 156 static void gen_addr32(int r, Sym *sym, int c)
 157 {
 158     if (r & VT_SYM)
 159         greloc(cur_text_section, sym, ind, R_386_32);
 160     gen_le32(c);
 161 }
 162
 163 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 164    opcode bits */
 165 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 166 {
 167     op_reg = op_reg << 3;
 168     if ((r & VT_VALMASK) == VT_CONST) {
 169         /* constant memory reference */
 170         o(0x05 | op_reg);
 171         gen_addr32(r, sym, c);
 172     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 173         /* currently, we use only ebp as base */
 174         if (c == (char)c) {
 175             /* short reference */
 176             o(0x45 | op_reg);
 177             g(c);
 178         } else {
 179             oad(0x85 | op_reg, c);
 180         }
 181     } else {
 182         g(0x00 | op_reg | (r & VT_VALMASK));
 183     }
 184 }
 185
 186
 187 /* load 'r' from value 'sv' */
 188 void load(int r, SValue *sv)
 189 {
 190     int v, t, ft, fc, fr;
 191     SValue v1;
 192
 193     fr = sv->r;
 194     ft = sv->type.t;
 195     fc = sv->c.ul;
 196
 197     v = fr & VT_VALMASK;
 198     if (fr & VT_LVAL) {
 199         if (v == VT_LLOCAL) {
 200             v1.type.t = VT_INT;
 201             v1.r = VT_LOCAL | VT_LVAL;
 202             v1.c.ul = fc;
 203             load(r, &v1);
 204             fr = r;
 205         }
 206         if ((ft & VT_BTYPE) == VT_FLOAT) {
 207             o(0xd9); /* flds */
 208             r = 0;
 209         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 210             o(0xdd); /* fldl */
 211             r = 0;
 212         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 213             o(0xdb); /* fldt */
 214             r = 5;
 215         } else if ((ft & VT_TYPE) == VT_BYTE) {
 216             o(0xbe0f);   /* movsbl */
 217         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 218             o(0xb60f);   /* movzbl */
 219         } else if ((ft & VT_TYPE) == VT_SHORT) {
 220             o(0xbf0f);   /* movswl */
 221         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 222             o(0xb70f);   /* movzwl */
 223         } else {
 224             o(0x8b);     /* movl */
 225         }
 226         gen_modrm(r, fr, sv->sym, fc);
 227     } else {
 228         if (v == VT_CONST) {
 229             o(0xb8 + r); /* mov $xx, r */
 230             gen_addr32(fr, sv->sym, fc);
 231         } else if (v == VT_LOCAL) {
 232             o(0x8d); /* lea xxx(%ebp), r */
 233             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 234         } else if (v == VT_CMP) {
 235             oad(0xb8 + r, 0); /* mov $0, r */
 236             o(0x0f); /* setxx %br */
 237             o(fc);
 238             o(0xc0 + r);
 239         } else if (v == VT_JMP || v == VT_JMPI) {
 240             t = v & 1;
 241             oad(0xb8 + r, t); /* mov $1, r */
 242             o(0x05eb); /* jmp after */
 243             gsym(fc);
 244             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 245         } else if (v != r) {
 246             o(0x89);
 247             o(0xc0 + r + v * 8); /* mov v, r */
 248         }
 249     }
 250 }
 251
 252 /* store register 'r' in lvalue 'v' */
 253 void store(int r, SValue *v)
 254 {
 255     int fr, bt, ft, fc;
 256
 257     ft = v->type.t;
 258     fc = v->c.ul;
 259     fr = v->r & VT_VALMASK;
 260     bt = ft & VT_BTYPE;
 261     /* XXX: incorrect if float reg to reg */
 262     if (bt == VT_FLOAT) {
 263         o(0xd9); /* fsts */
 264         r = 2;
 265     } else if (bt == VT_DOUBLE) {
 266         o(0xdd); /* fstpl */
 267         r = 2;
 268     } else if (bt == VT_LDOUBLE) {
 269         o(0xc0d9); /* fld %st(0) */
 270         o(0xdb); /* fstpt */
 271         r = 7;
 272     } else {
 273         if (bt == VT_SHORT)
 274             o(0x66);
 275         if (bt == VT_BYTE || bt == VT_BOOL)
 276             o(0x88);
 277         else
 278             o(0x89);
 279     }
 280     if (fr == VT_CONST ||
 281         fr == VT_LOCAL ||
 282         (v->r & VT_LVAL)) {
 283         gen_modrm(r, v->r, v->sym, fc);
 284     } else if (fr != r) {
 285         o(0xc0 + fr + r * 8); /* mov r, fr */
 286     }
 287 }
 288
 289 static void gadd_sp(int val)
 290 {
 291     if (val == (char)val) {
 292         o(0xc483);
 293         g(val);
 294     } else {
 295         oad(0xc481, val); /* add $xxx, %esp */
 296     }
 297 }
 298
 299 /* 'is_jmp' is '1' if it is a jump */
 300 static void gcall_or_jmp(int is_jmp)
 301 {
 302     int r;
 303     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 304         /* constant case */
 305         if (vtop->r & VT_SYM) {
 306             /* relocation case */
 307             greloc(cur_text_section, vtop->sym,
 308                    ind + 1, R_386_PC32);
 309         } else {
 310             /* put an empty PC32 relocation */
 311             put_elf_reloc(symtab_section, cur_text_section,
 312                           ind + 1, R_386_PC32, 0);
 313         }
 314         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 315     } else {
 316         /* otherwise, indirect call */
 317         r = gv(RC_INT);
 318         o(0xff); /* call/jmp *r */
 319         o(0xd0 + r + (is_jmp << 4));
 320     }
 321 }
 322
 323 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 324
 325 /* Generate function call. The function address is pushed first, then
 326    all the parameters in call order. This functions pops all the
 327    parameters and the function address. */
 328 void gfunc_call(int nb_args)
 329 {
 330     int size, align, r, args_size, i, func_call;
 331     Sym *func_sym;
 332
 333     args_size = 0;
 334     for(i = 0;i < nb_args; i++) {
 335         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 336             size = type_size(&vtop->type, &align);
 337             /* align to stack align size */
 338             size = (size + 3) & ~3;
 339             /* allocate the necessary size on stack */
 340             oad(0xec81, size); /* sub $xxx, %esp */
 341             /* generate structure store */
 342             r = get_reg(RC_INT);
 343             o(0x89); /* mov %esp, r */
 344             o(0xe0 + r);
 345             vset(&vtop->type, r | VT_LVAL, 0);
 346             vswap();
 347             vstore();
 348             args_size += size;
 349         } else if (is_float(vtop->type.t)) {
 350             gv(RC_FLOAT); /* only one float register */
 351             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 352                 size = 4;
 353             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 354                 size = 8;
 355             else
 356                 size = 12;
 357             oad(0xec81, size); /* sub $xxx, %esp */
 358             if (size == 12)
 359                 o(0x7cdb);
 360             else
 361                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 362             g(0x24);
 363             g(0x00);
 364             args_size += size;
 365         } else {
 366             /* simple type (currently always same size) */
 367             /* XXX: implicit cast ? */
 368             r = gv(RC_INT);
 369             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 370                 size = 8;
 371                 o(0x50 + vtop->r2); /* push r */
 372             } else {
 373                 size = 4;
 374             }
 375             o(0x50 + r); /* push r */
 376             args_size += size;
 377         }
 378         vtop--;
 379     }
 380     save_regs(0); /* save used temporary registers */
 381     func_sym = vtop->type.ref;
 382     func_call = func_sym->r;
 383     /* fast call case */
 384     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 385         int fastcall_nb_regs;
 386         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 387         for(i = 0;i < fastcall_nb_regs; i++) {
 388             if (args_size <= 0)
 389                 break;
 390             o(0x58 + fastcall_regs[i]); /* pop r */
 391             /* XXX: incorrect for struct/floats */
 392             args_size -= 4;
 393         }
 394     }
 395     gcall_or_jmp(0);
 396     if (args_size && func_sym->r != FUNC_STDCALL)
 397         gadd_sp(args_size);
 398     vtop--;
 399 }
 400
 401 #ifdef TCC_TARGET_PE
 402 #define FUNC_PROLOG_SIZE 10
 403 #else
 404 #define FUNC_PROLOG_SIZE 9
 405 #endif
 406
 407 /* generate function prolog of type 't' */
 408 void gfunc_prolog(CType *func_type)
 409 {
 410     int addr, align, size, func_call, fastcall_nb_regs;
 411     int param_index, param_addr;
 412     Sym *sym;
 413     CType *type;
 414
 415     sym = func_type->ref;
 416     func_call = sym->r;
 417     addr = 8;
 418     loc = 0;
 419     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 420         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 421     } else {
 422         fastcall_nb_regs = 0;
 423     }
 424     param_index = 0;
 425
 426     ind += FUNC_PROLOG_SIZE;
 427     func_sub_sp_offset = ind;
 428     /* if the function returns a structure, then add an
 429        implicit pointer parameter */
 430     func_vt = sym->type;
 431     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 432         /* XXX: fastcall case ? */
 433         func_vc = addr;
 434         addr += 4;
 435         param_index++;
 436     }
 437     /* define parameters */
 438     while ((sym = sym->next) != NULL) {
 439         type = &sym->type;
 440         size = type_size(type, &align);
 441         size = (size + 3) & ~3;
 442 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 443         /* structs are passed as pointer */
 444         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 445             size = 4;
 446         }
 447 #endif
 448         if (param_index < fastcall_nb_regs) {
 449             /* save FASTCALL register */
 450             loc -= 4;
 451             o(0x89);     /* movl */
 452             gen_modrm(fastcall_regs[param_index], VT_LOCAL, NULL, loc);
 453             param_addr = loc;
 454         } else {
 455             param_addr = addr;
 456             addr += size;
 457         }
 458         sym_push(sym->v & ~SYM_FIELD, type,
 459                  VT_LOCAL | VT_LVAL, param_addr);
 460         param_index++;
 461     }
 462     func_ret_sub = 0;
 463     /* pascal type call ? */
 464     if (func_call == FUNC_STDCALL)
 465         func_ret_sub = addr - 8;
 466
 467     /* leave some room for bound checking code */
 468     if (do_bounds_check) {
 469         oad(0xb8, 0); /* lbound section pointer */
 470         oad(0xb8, 0); /* call to function */
 471         func_bound_offset = lbounds_section->data_offset;
 472     }
 473 }
 474
 475 /* generate function epilog */
 476 void gfunc_epilog(void)
 477 {
 478     int v, saved_ind;
 479
 480 #ifdef CONFIG_TCC_BCHECK
 481     if (do_bounds_check && func_bound_offset != lbounds_section->data_offset) {
 482         int saved_ind;
 483         int *bounds_ptr;
 484         Sym *sym, *sym_data;
 485         /* add end of table info */
 486         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 487         *bounds_ptr = 0;
 488         /* generate bound local allocation */
 489         saved_ind = ind;
 490         ind = func_sub_sp_offset;
 491         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 492                                func_bound_offset, lbounds_section->data_offset);
 493         greloc(cur_text_section, sym_data,
 494                ind + 1, R_386_32);
 495         oad(0xb8, 0); /* mov %eax, xxx */
 496         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 497         greloc(cur_text_section, sym,
 498                ind + 1, R_386_PC32);
 499         oad(0xe8, -4);
 500         ind = saved_ind;
 501         /* generate bound check local freeing */
 502         o(0x5250); /* save returned value, if any */
 503         greloc(cur_text_section, sym_data,
 504                ind + 1, R_386_32);
 505         oad(0xb8, 0); /* mov %eax, xxx */
 506         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 507         greloc(cur_text_section, sym,
 508                ind + 1, R_386_PC32);
 509         oad(0xe8, -4);
 510         o(0x585a); /* restore returned value, if any */
 511     }
 512 #endif
 513     o(0xc9); /* leave */
 514     if (func_ret_sub == 0) {
 515         o(0xc3); /* ret */
 516     } else {
 517         o(0xc2); /* ret n */
 518         g(func_ret_sub);
 519         g(func_ret_sub >> 8);
 520     }
 521     /* align local size to word & save local variables */
 522
 523     v = (-loc + 3) & -4;
 524     saved_ind = ind;
 525     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 526 #ifdef TCC_TARGET_PE
 527     if (v >= 4096) {
 528         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 529         oad(0xb8, v); /* mov stacksize, %eax */
 530         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 531         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 532     } else
 533 #endif
 534     {
 535         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 536         o(0xec81);  /* sub esp, stacksize */
 537         gen_le32(v);
 538 #if FUNC_PROLOG_SIZE == 10
 539         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 540 #endif
 541     }
 542     ind = saved_ind;
 543 }
 544
 545 /* generate a jump to a label */
 546 int gjmp(int t)
 547 {
 548     return psym(0xe9, t);
 549 }
 550
 551 /* generate a jump to a fixed address */
 552 void gjmp_addr(int a)
 553 {
 554     int r;
 555     r = a - ind - 2;
 556     if (r == (char)r) {
 557         g(0xeb);
 558         g(r);
 559     } else {
 560         oad(0xe9, a - ind - 5);
 561     }
 562 }
 563
 564 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 565 int gtst(int inv, int t)
 566 {
 567     int v, *p;
 568
 569     v = vtop->r & VT_VALMASK;
 570     if (v == VT_CMP) {
 571         /* fast case : can jump directly since flags are set */
 572         g(0x0f);
 573         t = psym((vtop->c.i - 16) ^ inv, t);
 574     } else if (v == VT_JMP || v == VT_JMPI) {
 575         /* && or || optimization */
 576         if ((v & 1) == inv) {
 577             /* insert vtop->c jump list in t */
 578             p = &vtop->c.i;
 579             while (*p != 0)
 580                 p = (int *)(cur_text_section->data + *p);
 581             *p = t;
 582             t = vtop->c.i;
 583         } else {
 584             t = gjmp(t);
 585             gsym(vtop->c.i);
 586         }
 587     } else {
 588         if (is_float(vtop->type.t)) {
 589             vpushi(0);
 590             gen_op(TOK_NE);
 591         }
 592         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 593             /* constant jmp optimization */
 594             if ((vtop->c.i != 0) != inv)
 595                 t = gjmp(t);
 596         } else {
 597             v = gv(RC_INT);
 598             o(0x85);
 599             o(0xc0 + v * 9);
 600             g(0x0f);
 601             t = psym(0x85 ^ inv, t);
 602         }
 603     }
 604     vtop--;
 605     return t;
 606 }
 607
 608 /* generate an integer binary operation */
 609 void gen_opi(int op)
 610 {
 611     int r, fr, opc, c;
 612
 613     switch(op) {
 614     case '+':
 615     case TOK_ADDC1: /* add with carry generation */
 616         opc = 0;
 617     gen_op8:
 618         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 619             /* constant case */
 620             vswap();
 621             r = gv(RC_INT);
 622             vswap();
 623             c = vtop->c.i;
 624             if (c == (char)c) {
 625                 /* XXX: generate inc and dec for smaller code ? */
 626                 o(0x83);
 627                 o(0xc0 | (opc << 3) | r);
 628                 g(c);
 629             } else {
 630                 o(0x81);
 631                 oad(0xc0 | (opc << 3) | r, c);
 632             }
 633         } else {
 634             gv2(RC_INT, RC_INT);
 635             r = vtop[-1].r;
 636             fr = vtop[0].r;
 637             o((opc << 3) | 0x01);
 638             o(0xc0 + r + fr * 8);
 639         }
 640         vtop--;
 641         if (op >= TOK_ULT && op <= TOK_GT) {
 642             vtop->r = VT_CMP;
 643             vtop->c.i = op;
 644         }
 645         break;
 646     case '-':
 647     case TOK_SUBC1: /* sub with carry generation */
 648         opc = 5;
 649         goto gen_op8;
 650     case TOK_ADDC2: /* add with carry use */
 651         opc = 2;
 652         goto gen_op8;
 653     case TOK_SUBC2: /* sub with carry use */
 654         opc = 3;
 655         goto gen_op8;
 656     case '&':
 657         opc = 4;
 658         goto gen_op8;
 659     case '^':
 660         opc = 6;
 661         goto gen_op8;
 662     case '|':
 663         opc = 1;
 664         goto gen_op8;
 665     case '*':
 666         gv2(RC_INT, RC_INT);
 667         r = vtop[-1].r;
 668         fr = vtop[0].r;
 669         vtop--;
 670         o(0xaf0f); /* imul fr, r */
 671         o(0xc0 + fr + r * 8);
 672         break;
 673     case TOK_SHL:
 674         opc = 4;
 675         goto gen_shift;
 676     case TOK_SHR:
 677         opc = 5;
 678         goto gen_shift;
 679     case TOK_SAR:
 680         opc = 7;
 681     gen_shift:
 682         opc = 0xc0 | (opc << 3);
 683         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 684             /* constant case */
 685             vswap();
 686             r = gv(RC_INT);
 687             vswap();
 688             c = vtop->c.i & 0x1f;
 689             o(0xc1); /* shl/shr/sar $xxx, r */
 690             o(opc | r);
 691             g(c);
 692         } else {
 693             /* we generate the shift in ecx */
 694             gv2(RC_INT, RC_ECX);
 695             r = vtop[-1].r;
 696             o(0xd3); /* shl/shr/sar %cl, r */
 697             o(opc | r);
 698         }
 699         vtop--;
 700         break;
 701     case '/':
 702     case TOK_UDIV:
 703     case TOK_PDIV:
 704     case '%':
 705     case TOK_UMOD:
 706     case TOK_UMULL:
 707         /* first operand must be in eax */
 708         /* XXX: need better constraint for second operand */
 709         gv2(RC_EAX, RC_ECX);
 710         r = vtop[-1].r;
 711         fr = vtop[0].r;
 712         vtop--;
 713         save_reg(TREG_EDX);
 714         if (op == TOK_UMULL) {
 715             o(0xf7); /* mul fr */
 716             o(0xe0 + fr);
 717             vtop->r2 = TREG_EDX;
 718             r = TREG_EAX;
 719         } else {
 720             if (op == TOK_UDIV || op == TOK_UMOD) {
 721                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 722                 o(0xf0 + fr);
 723             } else {
 724                 o(0xf799); /* cltd, idiv fr, %eax */
 725                 o(0xf8 + fr);
 726             }
 727             if (op == '%' || op == TOK_UMOD)
 728                 r = TREG_EDX;
 729             else
 730                 r = TREG_EAX;
 731         }
 732         vtop->r = r;
 733         break;
 734     default:
 735         opc = 7;
 736         goto gen_op8;
 737     }
 738 }
 739
 740 /* generate a floating point operation 'v = t1 op t2' instruction. The
 741    two operands are guaranted to have the same floating point type */
 742 /* XXX: need to use ST1 too */
 743 void gen_opf(int op)
 744 {
 745     int a, ft, fc, swapped, r;
 746
 747     /* convert constants to memory references */
 748     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 749         vswap();
 750         gv(RC_FLOAT);
 751         vswap();
 752     }
 753     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 754         gv(RC_FLOAT);
 755
 756     /* must put at least one value in the floating point register */
 757     if ((vtop[-1].r & VT_LVAL) &&
 758         (vtop[0].r & VT_LVAL)) {
 759         vswap();
 760         gv(RC_FLOAT);
 761         vswap();
 762     }
 763     swapped = 0;
 764     /* swap the stack if needed so that t1 is the register and t2 is
 765        the memory reference */
 766     if (vtop[-1].r & VT_LVAL) {
 767         vswap();
 768         swapped = 1;
 769     }
 770     if (op >= TOK_ULT && op <= TOK_GT) {
 771         /* load on stack second operand */
 772         load(TREG_ST0, vtop);
 773         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 774         if (op == TOK_GE || op == TOK_GT)
 775             swapped = !swapped;
 776         else if (op == TOK_EQ || op == TOK_NE)
 777             swapped = 0;
 778         if (swapped)
 779             o(0xc9d9); /* fxch %st(1) */
 780         o(0xe9da); /* fucompp */
 781         o(0xe0df); /* fnstsw %ax */
 782         if (op == TOK_EQ) {
 783             o(0x45e480); /* and $0x45, %ah */
 784             o(0x40fC80); /* cmp $0x40, %ah */
 785         } else if (op == TOK_NE) {
 786             o(0x45e480); /* and $0x45, %ah */
 787             o(0x40f480); /* xor $0x40, %ah */
 788             op = TOK_NE;
 789         } else if (op == TOK_GE || op == TOK_LE) {
 790             o(0x05c4f6); /* test $0x05, %ah */
 791             op = TOK_EQ;
 792         } else {
 793             o(0x45c4f6); /* test $0x45, %ah */
 794             op = TOK_EQ;
 795         }
 796         vtop--;
 797         vtop->r = VT_CMP;
 798         vtop->c.i = op;
 799     } else {
 800         /* no memory reference possible for long double operations */
 801         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 802             load(TREG_ST0, vtop);
 803             swapped = !swapped;
 804         }
 805
 806         switch(op) {
 807         default:
 808         case '+':
 809             a = 0;
 810             break;
 811         case '-':
 812             a = 4;
 813             if (swapped)
 814                 a++;
 815             break;
 816         case '*':
 817             a = 1;
 818             break;
 819         case '/':
 820             a = 6;
 821             if (swapped)
 822                 a++;
 823             break;
 824         }
 825         ft = vtop->type.t;
 826         fc = vtop->c.ul;
 827         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 828             o(0xde); /* fxxxp %st, %st(1) */
 829             o(0xc1 + (a << 3));
 830         } else {
 831             /* if saved lvalue, then we must reload it */
 832             r = vtop->r;
 833             if ((r & VT_VALMASK) == VT_LLOCAL) {
 834                 SValue v1;
 835                 r = get_reg(RC_INT);
 836                 v1.type.t = VT_INT;
 837                 v1.r = VT_LOCAL | VT_LVAL;
 838                 v1.c.ul = fc;
 839                 load(r, &v1);
 840                 fc = 0;
 841             }
 842
 843             if ((ft & VT_BTYPE) == VT_DOUBLE)
 844                 o(0xdc);
 845             else
 846                 o(0xd8);
 847             gen_modrm(a, r, vtop->sym, fc);
 848         }
 849         vtop--;
 850     }
 851 }
 852
 853 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 854    and 'long long' cases. */
 855 void gen_cvt_itof(int t)
 856 {
 857     save_reg(TREG_ST0);
 858     gv(RC_INT);
 859     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 860         /* signed long long to float/double/long double (unsigned case
 861            is handled generically) */
 862         o(0x50 + vtop->r2); /* push r2 */
 863         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 864         o(0x242cdf); /* fildll (%esp) */
 865         o(0x08c483); /* add $8, %esp */
 866     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 867                (VT_INT | VT_UNSIGNED)) {
 868         /* unsigned int to float/double/long double */
 869         o(0x6a); /* push $0 */
 870         g(0x00);
 871         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 872         o(0x242cdf); /* fildll (%esp) */
 873         o(0x08c483); /* add $8, %esp */
 874     } else {
 875         /* int to float/double/long double */
 876         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 877         o(0x2404db); /* fildl (%esp) */
 878         o(0x04c483); /* add $4, %esp */
 879     }
 880     vtop->r = TREG_ST0;
 881 }
 882
 883 /* convert fp to int 't' type */
 884 /* XXX: handle long long case */
 885 void gen_cvt_ftoi(int t)
 886 {
 887     int r, r2, size;
 888     Sym *sym;
 889     CType ushort_type;
 890
 891     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 892
 893     gv(RC_FLOAT);
 894     if (t != VT_INT)
 895         size = 8;
 896     else
 897         size = 4;
 898
 899     o(0x2dd9); /* ldcw xxx */
 900     sym = external_global_sym(TOK___tcc_int_fpu_control,
 901                               &ushort_type, VT_LVAL);
 902     greloc(cur_text_section, sym,
 903            ind, R_386_32);
 904     gen_le32(0);
 905
 906     oad(0xec81, size); /* sub $xxx, %esp */
 907     if (size == 4)
 908         o(0x1cdb); /* fistpl */
 909     else
 910         o(0x3cdf); /* fistpll */
 911     o(0x24);
 912     o(0x2dd9); /* ldcw xxx */
 913     sym = external_global_sym(TOK___tcc_fpu_control,
 914                               &ushort_type, VT_LVAL);
 915     greloc(cur_text_section, sym,
 916            ind, R_386_32);
 917     gen_le32(0);
 918
 919     r = get_reg(RC_INT);
 920     o(0x58 + r); /* pop r */
 921     if (size == 8) {
 922         if (t == VT_LLONG) {
 923             vtop->r = r; /* mark reg as used */
 924             r2 = get_reg(RC_INT);
 925             o(0x58 + r2); /* pop r2 */
 926             vtop->r2 = r2;
 927         } else {
 928             o(0x04c483); /* add $4, %esp */
 929         }
 930     }
 931     vtop->r = r;
 932 }
 933
 934 /* convert from one floating point type to another */
 935 void gen_cvt_ftof(int t)
 936 {
 937     /* all we have to do on i386 is to put the float in a register */
 938     gv(RC_FLOAT);
 939 }
 940
 941 /* computed goto support */
 942 void ggoto(void)
 943 {
 944     gcall_or_jmp(1);
 945     vtop--;
 946 }
 947
 948 /* bound check support functions */
 949 #ifdef CONFIG_TCC_BCHECK
 950
 951 /* generate a bounded pointer addition */
 952 void gen_bounded_ptr_add(void)
 953 {
 954     Sym *sym;
 955
 956     /* prepare fast i386 function call (args in eax and edx) */
 957     gv2(RC_EAX, RC_EDX);
 958     /* save all temporary registers */
 959     vtop -= 2;
 960     save_regs(0);
 961     /* do a fast function call */
 962     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
 963     greloc(cur_text_section, sym,
 964            ind + 1, R_386_PC32);
 965     oad(0xe8, -4);
 966     /* returned pointer is in eax */
 967     vtop++;
 968     vtop->r = TREG_EAX | VT_BOUNDED;
 969     /* address of bounding function call point */
 970     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 971 }
 972
 973 /* patch pointer addition in vtop so that pointer dereferencing is
 974    also tested */
 975 void gen_bounded_ptr_deref(void)
 976 {
 977     int func;
 978     int size, align;
 979     Elf32_Rel *rel;
 980     Sym *sym;
 981
 982     size = 0;
 983     /* XXX: put that code in generic part of tcc */
 984     if (!is_float(vtop->type.t)) {
 985         if (vtop->r & VT_LVAL_BYTE)
 986             size = 1;
 987         else if (vtop->r & VT_LVAL_SHORT)
 988             size = 2;
 989     }
 990     if (!size)
 991         size = type_size(&vtop->type, &align);
 992     switch(size) {
 993     case  1: func = TOK___bound_ptr_indir1; break;
 994     case  2: func = TOK___bound_ptr_indir2; break;
 995     case  4: func = TOK___bound_ptr_indir4; break;
 996     case  8: func = TOK___bound_ptr_indir8; break;
 997     case 12: func = TOK___bound_ptr_indir12; break;
 998     case 16: func = TOK___bound_ptr_indir16; break;
 999     default:
1000         error("unhandled size when derefencing bounded pointer");
1001         func = 0;
1002         break;
1003     }
1004
1005     /* patch relocation */
1006     /* XXX: find a better solution ? */
1007     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1008     sym = external_global_sym(func, &func_old_type, 0);
1009     if (!sym->c)
1010         put_extern_sym(sym, NULL, 0, 0);
1011     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1012 }
1013 #endif
1014
1015 /* end of X86 code generator */
1016 /*************************************************************/
1017