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 #include "tcc.h"
  22
  23 const int reg_classes[NB_REGS] = {
  24     /* eax */ RC_INT | RC_EAX,
  25     /* ecx */ RC_INT | RC_ECX,
  26     /* edx */ RC_INT | RC_EDX,
  27     /* st0 */ RC_FLOAT | RC_ST0,
  28 };
  29
  30
  31 /******************************************************/
  32
  33 static unsigned long func_sub_sp_offset;
  34 static unsigned long func_bound_offset;
  35 static int func_ret_sub;
  36
  37 /* XXX: make it faster ? */
  38 void g(int c)
  39 {
  40     int ind1;
  41     ind1 = ind + 1;
  42     if (ind1 > cur_text_section->data_allocated)
  43         section_realloc(cur_text_section, ind1);
  44     cur_text_section->data[ind] = c;
  45     ind = ind1;
  46 }
  47
  48 void o(unsigned int c)
  49 {
  50     while (c) {
  51         g(c);
  52         c = c >> 8;
  53     }
  54 }
  55
  56 void gen_le32(int c)
  57 {
  58     g(c);
  59     g(c >> 8);
  60     g(c >> 16);
  61     g(c >> 24);
  62 }
  63
  64 /* output a symbol and patch all calls to it */
  65 void gsym_addr(int t, int a)
  66 {
  67     int n, *ptr;
  68     while (t) {
  69         ptr = (int *)(cur_text_section->data + t);
  70         n = *ptr; /* next value */
  71         *ptr = a - t - 4;
  72         t = n;
  73     }
  74 }
  75
  76 void gsym(int t)
  77 {
  78     gsym_addr(t, ind);
  79 }
  80
  81 /* psym is used to put an instruction with a data field which is a
  82    reference to a symbol. It is in fact the same as oad ! */
  83 #define psym oad
  84
  85 /* instruction + 4 bytes data. Return the address of the data */
  86 int oad(int c, int s)
  87 {
  88     int ind1;
  89
  90     o(c);
  91     ind1 = ind + 4;
  92     if (ind1 > cur_text_section->data_allocated)
  93         section_realloc(cur_text_section, ind1);
  94     *(int *)(cur_text_section->data + ind) = s;
  95     s = ind;
  96     ind = ind1;
  97     return s;
  98 }
  99
 100 /* output constant with relocation if 'r & VT_SYM' is true */
 101 static void gen_addr32(int r, Sym *sym, int c)
 102 {
 103     if (r & VT_SYM)
 104         greloc(cur_text_section, sym, ind, R_386_32);
 105     gen_le32(c);
 106 }
 107
 108 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 109    opcode bits */
 110 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 111 {
 112     op_reg = op_reg << 3;
 113     if ((r & VT_VALMASK) == VT_CONST) {
 114         /* constant memory reference */
 115         o(0x05 | op_reg);
 116         gen_addr32(r, sym, c);
 117     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 118         /* currently, we use only ebp as base */
 119         if (c == (char)c) {
 120             /* short reference */
 121             o(0x45 | op_reg);
 122             g(c);
 123         } else {
 124             oad(0x85 | op_reg, c);
 125         }
 126     } else {
 127         g(0x00 | op_reg | (r & VT_VALMASK));
 128     }
 129 }
 130
 131 #ifdef TCC_TARGET_PE
 132 static void indir(void);
 133
 134 int handle_dllimport(int r, SValue *sv, void (*fn)(int r, SValue *sv))
 135 {
 136     if ((sv->r & (VT_VALMASK|VT_SYM|VT_CONST)) != (VT_SYM|VT_CONST))
 137         return 0;
 138     if (0 == (sv->sym->type.t & VT_IMPORT))
 139         return 0;
 140
 141     printf("import %d %04x %s\n", r, ind, get_tok_str(sv->sym->v, NULL));
 142
 143     sv->sym->type.t &= ~VT_IMPORT;
 144     ++vtop;
 145
 146     *vtop = *sv;
 147     mk_pointer(&vtop->type);
 148     indir();
 149     fn(r, vtop);
 150
 151     --vtop;
 152     sv->sym->type.t |= VT_IMPORT;
 153     return 1;
 154 }
 155 #endif
 156
 157 /* load 'r' from value 'sv' */
 158 void load(int r, SValue *sv)
 159 {
 160     int v, t, ft, fc, fr;
 161     SValue v1;
 162
 163 #ifdef TCC_TARGET_PE
 164     if (handle_dllimport(r, sv, load))
 165         return;
 166 #endif
 167     fr = sv->r;
 168     ft = sv->type.t;
 169     fc = sv->c.ul;
 170
 171     v = fr & VT_VALMASK;
 172     if (fr & VT_LVAL) {
 173         if (v == VT_LLOCAL) {
 174             v1.type.t = VT_INT;
 175             v1.r = VT_LOCAL | VT_LVAL;
 176             v1.c.ul = fc;
 177             load(r, &v1);
 178             fr = r;
 179         }
 180         if ((ft & VT_BTYPE) == VT_FLOAT) {
 181             o(0xd9); /* flds */
 182             r = 0;
 183         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 184             o(0xdd); /* fldl */
 185             r = 0;
 186         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 187             o(0xdb); /* fldt */
 188             r = 5;
 189         } else if ((ft & VT_TYPE) == VT_BYTE) {
 190             o(0xbe0f);   /* movsbl */
 191         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 192             o(0xb60f);   /* movzbl */
 193         } else if ((ft & VT_TYPE) == VT_SHORT) {
 194             o(0xbf0f);   /* movswl */
 195         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 196             o(0xb70f);   /* movzwl */
 197         } else {
 198             o(0x8b);     /* movl */
 199         }
 200         gen_modrm(r, fr, sv->sym, fc);
 201     } else {
 202         if (v == VT_CONST) {
 203             o(0xb8 + r); /* mov $xx, r */
 204             gen_addr32(fr, sv->sym, fc);
 205         } else if (v == VT_LOCAL) {
 206             o(0x8d); /* lea xxx(%ebp), r */
 207             gen_modrm(r, VT_LOCAL, sv->sym, fc);
 208         } else if (v == VT_CMP) {
 209             oad(0xb8 + r, 0); /* mov $0, r */
 210             o(0x0f); /* setxx %br */
 211             o(fc);
 212             o(0xc0 + r);
 213         } else if (v == VT_JMP || v == VT_JMPI) {
 214             t = v & 1;
 215             oad(0xb8 + r, t); /* mov $1, r */
 216             o(0x05eb); /* jmp after */
 217             gsym(fc);
 218             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 219         } else if (v != r) {
 220             o(0x89);
 221             o(0xc0 + r + v * 8); /* mov v, r */
 222         }
 223     }
 224 }
 225
 226 /* store register 'r' in lvalue 'v' */
 227 void store(int r, SValue *v)
 228 {
 229     int fr, bt, ft, fc;
 230
 231 #ifdef TCC_TARGET_PE
 232     if (handle_dllimport(r, v, store))
 233         return;
 234 #endif
 235     ft = v->type.t;
 236     fc = v->c.ul;
 237     fr = v->r & VT_VALMASK;
 238     bt = ft & VT_BTYPE;
 239     /* XXX: incorrect if float reg to reg */
 240     if (bt == VT_FLOAT) {
 241         o(0xd9); /* fsts */
 242         r = 2;
 243     } else if (bt == VT_DOUBLE) {
 244         o(0xdd); /* fstpl */
 245         r = 2;
 246     } else if (bt == VT_LDOUBLE) {
 247         o(0xc0d9); /* fld %st(0) */
 248         o(0xdb); /* fstpt */
 249         r = 7;
 250     } else {
 251         if (bt == VT_SHORT)
 252             o(0x66);
 253         if (bt == VT_BYTE || bt == VT_BOOL)
 254             o(0x88);
 255         else
 256             o(0x89);
 257     }
 258     if (fr == VT_CONST ||
 259         fr == VT_LOCAL ||
 260         (v->r & VT_LVAL)) {
 261         gen_modrm(r, v->r, v->sym, fc);
 262     } else if (fr != r) {
 263         o(0xc0 + fr + r * 8); /* mov r, fr */
 264     }
 265 }
 266
 267 static void gadd_sp(int val)
 268 {
 269     if (val == (char)val) {
 270         o(0xc483);
 271         g(val);
 272     } else {
 273         oad(0xc481, val); /* add $xxx, %esp */
 274     }
 275 }
 276
 277 /* 'is_jmp' is '1' if it is a jump */
 278 static void gcall_or_jmp(int is_jmp)
 279 {
 280     int r;
 281     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 282         /* constant case */
 283         if (vtop->r & VT_SYM) {
 284             /* relocation case */
 285             greloc(cur_text_section, vtop->sym,
 286                    ind + 1, R_386_PC32);
 287         } else {
 288             /* put an empty PC32 relocation */
 289             put_elf_reloc(symtab_section, cur_text_section,
 290                           ind + 1, R_386_PC32, 0);
 291         }
 292         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 293     } else {
 294         /* otherwise, indirect call */
 295         r = gv(RC_INT);
 296         o(0xff); /* call/jmp *r */
 297         o(0xd0 + r + (is_jmp << 4));
 298     }
 299 }
 300
 301 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 302 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 303
 304 /* Generate function call. The function address is pushed first, then
 305    all the parameters in call order. This functions pops all the
 306    parameters and the function address. */
 307 void gfunc_call(int nb_args)
 308 {
 309     int size, align, r, args_size, i, func_call;
 310     Sym *func_sym;
 311
 312     args_size = 0;
 313     for(i = 0;i < nb_args; i++) {
 314         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 315             size = type_size(&vtop->type, &align);
 316             /* align to stack align size */
 317             size = (size + 3) & ~3;
 318             /* allocate the necessary size on stack */
 319             oad(0xec81, size); /* sub $xxx, %esp */
 320             /* generate structure store */
 321             r = get_reg(RC_INT);
 322             o(0x89); /* mov %esp, r */
 323             o(0xe0 + r);
 324             vset(&vtop->type, r | VT_LVAL, 0);
 325             vswap();
 326             vstore();
 327             args_size += size;
 328         } else if (is_float(vtop->type.t)) {
 329             gv(RC_FLOAT); /* only one float register */
 330             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 331                 size = 4;
 332             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 333                 size = 8;
 334             else
 335                 size = 12;
 336             oad(0xec81, size); /* sub $xxx, %esp */
 337             if (size == 12)
 338                 o(0x7cdb);
 339             else
 340                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 341             g(0x24);
 342             g(0x00);
 343             args_size += size;
 344         } else {
 345             /* simple type (currently always same size) */
 346             /* XXX: implicit cast ? */
 347             r = gv(RC_INT);
 348             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 349                 size = 8;
 350                 o(0x50 + vtop->r2); /* push r */
 351             } else {
 352                 size = 4;
 353             }
 354             o(0x50 + r); /* push r */
 355             args_size += size;
 356         }
 357         vtop--;
 358     }
 359     save_regs(0); /* save used temporary registers */
 360     func_sym = vtop->type.ref;
 361     func_call = FUNC_CALL(func_sym->r);
 362     /* fast call case */
 363     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 364         func_call == FUNC_FASTCALLW) {
 365         int fastcall_nb_regs;
 366         uint8_t *fastcall_regs_ptr;
 367         if (func_call == FUNC_FASTCALLW) {
 368             fastcall_regs_ptr = fastcallw_regs;
 369             fastcall_nb_regs = 2;
 370         } else {
 371             fastcall_regs_ptr = fastcall_regs;
 372             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 373         }
 374         for(i = 0;i < fastcall_nb_regs; i++) {
 375             if (args_size <= 0)
 376                 break;
 377             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 378             /* XXX: incorrect for struct/floats */
 379             args_size -= 4;
 380         }
 381     }
 382     gcall_or_jmp(0);
 383
 384 #ifdef TCC_TARGET_PE
 385     if ((func_sym->type.t & VT_BTYPE) == VT_STRUCT)
 386         args_size -= 4;
 387 #endif
 388     if (args_size && func_call != FUNC_STDCALL)
 389         gadd_sp(args_size);
 390     vtop--;
 391 }
 392
 393 #ifdef TCC_TARGET_PE
 394 #define FUNC_PROLOG_SIZE 10
 395 #else
 396 #define FUNC_PROLOG_SIZE 9
 397 #endif
 398
 399 /* generate function prolog of type 't' */
 400 void gfunc_prolog(CType *func_type)
 401 {
 402     int addr, align, size, func_call, fastcall_nb_regs;
 403     int param_index, param_addr;
 404     uint8_t *fastcall_regs_ptr;
 405     Sym *sym;
 406     CType *type;
 407
 408     sym = func_type->ref;
 409     func_call = FUNC_CALL(sym->r);
 410     addr = 8;
 411     loc = 0;
 412     func_vc = 0;
 413
 414     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 415         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 416         fastcall_regs_ptr = fastcall_regs;
 417     } else if (func_call == FUNC_FASTCALLW) {
 418         fastcall_nb_regs = 2;
 419         fastcall_regs_ptr = fastcallw_regs;
 420     } else {
 421         fastcall_nb_regs = 0;
 422         fastcall_regs_ptr = NULL;
 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_ptr[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 | lvalue_type(type->t), 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 #ifdef TCC_TARGET_PE
 467     else if (func_vc)
 468         func_ret_sub = 4;
 469 #endif
 470
 471     /* leave some room for bound checking code */
 472     if (tcc_state->do_bounds_check) {
 473         oad(0xb8, 0); /* lbound section pointer */
 474         oad(0xb8, 0); /* call to function */
 475         func_bound_offset = lbounds_section->data_offset;
 476     }
 477 }
 478
 479 /* generate function epilog */
 480 void gfunc_epilog(void)
 481 {
 482     int v, saved_ind;
 483
 484 #ifdef CONFIG_TCC_BCHECK
 485     if (tcc_state->do_bounds_check
 486      && func_bound_offset != lbounds_section->data_offset) {
 487         int saved_ind;
 488         int *bounds_ptr;
 489         Sym *sym, *sym_data;
 490         /* add end of table info */
 491         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 492         *bounds_ptr = 0;
 493         /* generate bound local allocation */
 494         saved_ind = ind;
 495         ind = func_sub_sp_offset;
 496         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 497                                func_bound_offset, lbounds_section->data_offset);
 498         greloc(cur_text_section, sym_data,
 499                ind + 1, R_386_32);
 500         oad(0xb8, 0); /* mov %eax, xxx */
 501         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 502         greloc(cur_text_section, sym,
 503                ind + 1, R_386_PC32);
 504         oad(0xe8, -4);
 505         ind = saved_ind;
 506         /* generate bound check local freeing */
 507         o(0x5250); /* save returned value, if any */
 508         greloc(cur_text_section, sym_data,
 509                ind + 1, R_386_32);
 510         oad(0xb8, 0); /* mov %eax, xxx */
 511         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 512         greloc(cur_text_section, sym,
 513                ind + 1, R_386_PC32);
 514         oad(0xe8, -4);
 515         o(0x585a); /* restore returned value, if any */
 516     }
 517 #endif
 518     o(0xc9); /* leave */
 519     if (func_ret_sub == 0) {
 520         o(0xc3); /* ret */
 521     } else {
 522         o(0xc2); /* ret n */
 523         g(func_ret_sub);
 524         g(func_ret_sub >> 8);
 525     }
 526     /* align local size to word & save local variables */
 527
 528     v = (-loc + 3) & -4;
 529     saved_ind = ind;
 530     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 531 #ifdef TCC_TARGET_PE
 532     if (v >= 4096) {
 533         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 534         oad(0xb8, v); /* mov stacksize, %eax */
 535         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 536         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 537     } else
 538 #endif
 539     {
 540         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 541         o(0xec81);  /* sub esp, stacksize */
 542         gen_le32(v);
 543 #if FUNC_PROLOG_SIZE == 10
 544         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 545 #endif
 546     }
 547     ind = saved_ind;
 548 }
 549
 550 /* generate a jump to a label */
 551 int gjmp(int t)
 552 {
 553     return psym(0xe9, t);
 554 }
 555
 556 /* generate a jump to a fixed address */
 557 void gjmp_addr(int a)
 558 {
 559     int r;
 560     r = a - ind - 2;
 561     if (r == (char)r) {
 562         g(0xeb);
 563         g(r);
 564     } else {
 565         oad(0xe9, a - ind - 5);
 566     }
 567 }
 568
 569 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 570 int gtst(int inv, int t)
 571 {
 572     int v, *p;
 573
 574     v = vtop->r & VT_VALMASK;
 575     if (v == VT_CMP) {
 576         /* fast case : can jump directly since flags are set */
 577         g(0x0f);
 578         t = psym((vtop->c.i - 16) ^ inv, t);
 579     } else if (v == VT_JMP || v == VT_JMPI) {
 580         /* && or || optimization */
 581         if ((v & 1) == inv) {
 582             /* insert vtop->c jump list in t */
 583             p = &vtop->c.i;
 584             while (*p != 0)
 585                 p = (int *)(cur_text_section->data + *p);
 586             *p = t;
 587             t = vtop->c.i;
 588         } else {
 589             t = gjmp(t);
 590             gsym(vtop->c.i);
 591         }
 592     } else {
 593         if (is_float(vtop->type.t) ||
 594             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 595             vpushi(0);
 596             gen_op(TOK_NE);
 597         }
 598         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 599             /* constant jmp optimization */
 600             if ((vtop->c.i != 0) != inv)
 601                 t = gjmp(t);
 602         } else {
 603             v = gv(RC_INT);
 604             o(0x85);
 605             o(0xc0 + v * 9);
 606             g(0x0f);
 607             t = psym(0x85 ^ inv, t);
 608         }
 609     }
 610     vtop--;
 611     return t;
 612 }
 613
 614 /* generate an integer binary operation */
 615 void gen_opi(int op)
 616 {
 617     int r, fr, opc, c;
 618
 619     switch(op) {
 620     case '+':
 621     case TOK_ADDC1: /* add with carry generation */
 622         opc = 0;
 623     gen_op8:
 624         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 625             /* constant case */
 626             vswap();
 627             r = gv(RC_INT);
 628             vswap();
 629             c = vtop->c.i;
 630             if (c == (char)c) {
 631                 /* XXX: generate inc and dec for smaller code ? */
 632                 o(0x83);
 633                 o(0xc0 | (opc << 3) | r);
 634                 g(c);
 635             } else {
 636                 o(0x81);
 637                 oad(0xc0 | (opc << 3) | r, c);
 638             }
 639         } else {
 640             gv2(RC_INT, RC_INT);
 641             r = vtop[-1].r;
 642             fr = vtop[0].r;
 643             o((opc << 3) | 0x01);
 644             o(0xc0 + r + fr * 8);
 645         }
 646         vtop--;
 647         if (op >= TOK_ULT && op <= TOK_GT) {
 648             vtop->r = VT_CMP;
 649             vtop->c.i = op;
 650         }
 651         break;
 652     case '-':
 653     case TOK_SUBC1: /* sub with carry generation */
 654         opc = 5;
 655         goto gen_op8;
 656     case TOK_ADDC2: /* add with carry use */
 657         opc = 2;
 658         goto gen_op8;
 659     case TOK_SUBC2: /* sub with carry use */
 660         opc = 3;
 661         goto gen_op8;
 662     case '&':
 663         opc = 4;
 664         goto gen_op8;
 665     case '^':
 666         opc = 6;
 667         goto gen_op8;
 668     case '|':
 669         opc = 1;
 670         goto gen_op8;
 671     case '*':
 672         gv2(RC_INT, RC_INT);
 673         r = vtop[-1].r;
 674         fr = vtop[0].r;
 675         vtop--;
 676         o(0xaf0f); /* imul fr, r */
 677         o(0xc0 + fr + r * 8);
 678         break;
 679     case TOK_SHL:
 680         opc = 4;
 681         goto gen_shift;
 682     case TOK_SHR:
 683         opc = 5;
 684         goto gen_shift;
 685     case TOK_SAR:
 686         opc = 7;
 687     gen_shift:
 688         opc = 0xc0 | (opc << 3);
 689         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 690             /* constant case */
 691             vswap();
 692             r = gv(RC_INT);
 693             vswap();
 694             c = vtop->c.i & 0x1f;
 695             o(0xc1); /* shl/shr/sar $xxx, r */
 696             o(opc | r);
 697             g(c);
 698         } else {
 699             /* we generate the shift in ecx */
 700             gv2(RC_INT, RC_ECX);
 701             r = vtop[-1].r;
 702             o(0xd3); /* shl/shr/sar %cl, r */
 703             o(opc | r);
 704         }
 705         vtop--;
 706         break;
 707     case '/':
 708     case TOK_UDIV:
 709     case TOK_PDIV:
 710     case '%':
 711     case TOK_UMOD:
 712     case TOK_UMULL:
 713         /* first operand must be in eax */
 714         /* XXX: need better constraint for second operand */
 715         gv2(RC_EAX, RC_ECX);
 716         r = vtop[-1].r;
 717         fr = vtop[0].r;
 718         vtop--;
 719         save_reg(TREG_EDX);
 720         if (op == TOK_UMULL) {
 721             o(0xf7); /* mul fr */
 722             o(0xe0 + fr);
 723             vtop->r2 = TREG_EDX;
 724             r = TREG_EAX;
 725         } else {
 726             if (op == TOK_UDIV || op == TOK_UMOD) {
 727                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 728                 o(0xf0 + fr);
 729             } else {
 730                 o(0xf799); /* cltd, idiv fr, %eax */
 731                 o(0xf8 + fr);
 732             }
 733             if (op == '%' || op == TOK_UMOD)
 734                 r = TREG_EDX;
 735             else
 736                 r = TREG_EAX;
 737         }
 738         vtop->r = r;
 739         break;
 740     default:
 741         opc = 7;
 742         goto gen_op8;
 743     }
 744 }
 745
 746 /* generate a floating point operation 'v = t1 op t2' instruction. The
 747    two operands are guaranted to have the same floating point type */
 748 /* XXX: need to use ST1 too */
 749 void gen_opf(int op)
 750 {
 751     int a, ft, fc, swapped, r;
 752
 753     /* convert constants to memory references */
 754     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 755         vswap();
 756         gv(RC_FLOAT);
 757         vswap();
 758     }
 759     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 760         gv(RC_FLOAT);
 761
 762     /* must put at least one value in the floating point register */
 763     if ((vtop[-1].r & VT_LVAL) &&
 764         (vtop[0].r & VT_LVAL)) {
 765         vswap();
 766         gv(RC_FLOAT);
 767         vswap();
 768     }
 769     swapped = 0;
 770     /* swap the stack if needed so that t1 is the register and t2 is
 771        the memory reference */
 772     if (vtop[-1].r & VT_LVAL) {
 773         vswap();
 774         swapped = 1;
 775     }
 776     if (op >= TOK_ULT && op <= TOK_GT) {
 777         /* load on stack second operand */
 778         load(TREG_ST0, vtop);
 779         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 780         if (op == TOK_GE || op == TOK_GT)
 781             swapped = !swapped;
 782         else if (op == TOK_EQ || op == TOK_NE)
 783             swapped = 0;
 784         if (swapped)
 785             o(0xc9d9); /* fxch %st(1) */
 786         o(0xe9da); /* fucompp */
 787         o(0xe0df); /* fnstsw %ax */
 788         if (op == TOK_EQ) {
 789             o(0x45e480); /* and $0x45, %ah */
 790             o(0x40fC80); /* cmp $0x40, %ah */
 791         } else if (op == TOK_NE) {
 792             o(0x45e480); /* and $0x45, %ah */
 793             o(0x40f480); /* xor $0x40, %ah */
 794             op = TOK_NE;
 795         } else if (op == TOK_GE || op == TOK_LE) {
 796             o(0x05c4f6); /* test $0x05, %ah */
 797             op = TOK_EQ;
 798         } else {
 799             o(0x45c4f6); /* test $0x45, %ah */
 800             op = TOK_EQ;
 801         }
 802         vtop--;
 803         vtop->r = VT_CMP;
 804         vtop->c.i = op;
 805     } else {
 806         /* no memory reference possible for long double operations */
 807         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 808             load(TREG_ST0, vtop);
 809             swapped = !swapped;
 810         }
 811
 812         switch(op) {
 813         default:
 814         case '+':
 815             a = 0;
 816             break;
 817         case '-':
 818             a = 4;
 819             if (swapped)
 820                 a++;
 821             break;
 822         case '*':
 823             a = 1;
 824             break;
 825         case '/':
 826             a = 6;
 827             if (swapped)
 828                 a++;
 829             break;
 830         }
 831         ft = vtop->type.t;
 832         fc = vtop->c.ul;
 833         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 834             o(0xde); /* fxxxp %st, %st(1) */
 835             o(0xc1 + (a << 3));
 836         } else {
 837             /* if saved lvalue, then we must reload it */
 838             r = vtop->r;
 839             if ((r & VT_VALMASK) == VT_LLOCAL) {
 840                 SValue v1;
 841                 r = get_reg(RC_INT);
 842                 v1.type.t = VT_INT;
 843                 v1.r = VT_LOCAL | VT_LVAL;
 844                 v1.c.ul = fc;
 845                 load(r, &v1);
 846                 fc = 0;
 847             }
 848
 849             if ((ft & VT_BTYPE) == VT_DOUBLE)
 850                 o(0xdc);
 851             else
 852                 o(0xd8);
 853             gen_modrm(a, r, vtop->sym, fc);
 854         }
 855         vtop--;
 856     }
 857 }
 858
 859 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 860    and 'long long' cases. */
 861 void gen_cvt_itof(int t)
 862 {
 863     save_reg(TREG_ST0);
 864     gv(RC_INT);
 865     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 866         /* signed long long to float/double/long double (unsigned case
 867            is handled generically) */
 868         o(0x50 + vtop->r2); /* push r2 */
 869         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 870         o(0x242cdf); /* fildll (%esp) */
 871         o(0x08c483); /* add $8, %esp */
 872     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 873                (VT_INT | VT_UNSIGNED)) {
 874         /* unsigned int to float/double/long double */
 875         o(0x6a); /* push $0 */
 876         g(0x00);
 877         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 878         o(0x242cdf); /* fildll (%esp) */
 879         o(0x08c483); /* add $8, %esp */
 880     } else {
 881         /* int to float/double/long double */
 882         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 883         o(0x2404db); /* fildl (%esp) */
 884         o(0x04c483); /* add $4, %esp */
 885     }
 886     vtop->r = TREG_ST0;
 887 }
 888
 889 /* convert fp to int 't' type */
 890 /* XXX: handle long long case */
 891 void gen_cvt_ftoi(int t)
 892 {
 893     int r, r2, size;
 894     Sym *sym;
 895     CType ushort_type;
 896
 897     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 898
 899     gv(RC_FLOAT);
 900     if (t != VT_INT)
 901         size = 8;
 902     else
 903         size = 4;
 904
 905     o(0x2dd9); /* ldcw xxx */
 906     sym = external_global_sym(TOK___tcc_int_fpu_control,
 907                               &ushort_type, VT_LVAL);
 908     greloc(cur_text_section, sym,
 909            ind, R_386_32);
 910     gen_le32(0);
 911
 912     oad(0xec81, size); /* sub $xxx, %esp */
 913     if (size == 4)
 914         o(0x1cdb); /* fistpl */
 915     else
 916         o(0x3cdf); /* fistpll */
 917     o(0x24);
 918     o(0x2dd9); /* ldcw xxx */
 919     sym = external_global_sym(TOK___tcc_fpu_control,
 920                               &ushort_type, VT_LVAL);
 921     greloc(cur_text_section, sym,
 922            ind, R_386_32);
 923     gen_le32(0);
 924
 925     r = get_reg(RC_INT);
 926     o(0x58 + r); /* pop r */
 927     if (size == 8) {
 928         if (t == VT_LLONG) {
 929             vtop->r = r; /* mark reg as used */
 930             r2 = get_reg(RC_INT);
 931             o(0x58 + r2); /* pop r2 */
 932             vtop->r2 = r2;
 933         } else {
 934             o(0x04c483); /* add $4, %esp */
 935         }
 936     }
 937     vtop->r = r;
 938 }
 939
 940 /* convert from one floating point type to another */
 941 void gen_cvt_ftof(int t)
 942 {
 943     /* all we have to do on i386 is to put the float in a register */
 944     gv(RC_FLOAT);
 945 }
 946
 947 /* computed goto support */
 948 void ggoto(void)
 949 {
 950     gcall_or_jmp(1);
 951     vtop--;
 952 }
 953
 954 /* bound check support functions */
 955 #ifdef CONFIG_TCC_BCHECK
 956
 957 /* generate a bounded pointer addition */
 958 void gen_bounded_ptr_add(void)
 959 {
 960     Sym *sym;
 961
 962     /* prepare fast i386 function call (args in eax and edx) */
 963     gv2(RC_EAX, RC_EDX);
 964     /* save all temporary registers */
 965     vtop -= 2;
 966     save_regs(0);
 967     /* do a fast function call */
 968     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
 969     greloc(cur_text_section, sym,
 970            ind + 1, R_386_PC32);
 971     oad(0xe8, -4);
 972     /* returned pointer is in eax */
 973     vtop++;
 974     vtop->r = TREG_EAX | VT_BOUNDED;
 975     /* address of bounding function call point */
 976     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
 977 }
 978
 979 /* patch pointer addition in vtop so that pointer dereferencing is
 980    also tested */
 981 void gen_bounded_ptr_deref(void)
 982 {
 983     int func;
 984     int size, align;
 985     Elf32_Rel *rel;
 986     Sym *sym;
 987
 988     size = 0;
 989     /* XXX: put that code in generic part of tcc */
 990     if (!is_float(vtop->type.t)) {
 991         if (vtop->r & VT_LVAL_BYTE)
 992             size = 1;
 993         else if (vtop->r & VT_LVAL_SHORT)
 994             size = 2;
 995     }
 996     if (!size)
 997         size = type_size(&vtop->type, &align);
 998     switch(size) {
 999     case  1: func = TOK___bound_ptr_indir1; break;
1000     case  2: func = TOK___bound_ptr_indir2; break;
1001     case  4: func = TOK___bound_ptr_indir4; break;
1002     case  8: func = TOK___bound_ptr_indir8; break;
1003     case 12: func = TOK___bound_ptr_indir12; break;
1004     case 16: func = TOK___bound_ptr_indir16; break;
1005     default:
1006         error("unhandled size when derefencing bounded pointer");
1007         func = 0;
1008         break;
1009     }
1010
1011     /* patch relocation */
1012     /* XXX: find a better solution ? */
1013     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1014     sym = external_global_sym(func, &func_old_type, 0);
1015     if (!sym->c)
1016         put_extern_sym(sym, NULL, 0, 0);
1017     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1018 }
1019 #endif
1020
1021 /* end of X86 code generator */
1022 /*************************************************************/
1023
1024
1025
1026
1027
1028
1029
1030
1031