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 #ifdef TARGET_DEFS_ONLY
  22
  23 /* number of available registers */
  24 #define NB_REGS         5
  25 #define NB_ASM_REGS     8
  26 #define CONFIG_TCC_ASM
  27
  28 /* a register can belong to several classes. The classes must be
  29    sorted from more general to more precise (see gv2() code which does
  30    assumptions on it). */
  31 #define RC_INT     0x0001 /* generic integer register */
  32 #define RC_FLOAT   0x0002 /* generic float register */
  33 #define RC_EAX     0x0004
  34 #define RC_ST0     0x0008
  35 #define RC_ECX     0x0010
  36 #define RC_EDX     0x0020
  37 #define RC_EBX     0x0040
  38
  39 #define RC_IRET    RC_EAX /* function return: integer register */
  40 #define RC_IRE2    RC_EDX /* function return: second integer register */
  41 #define RC_FRET    RC_ST0 /* function return: float register */
  42
  43 /* pretty names for the registers */
  44 enum {
  45     TREG_EAX = 0,
  46     TREG_ECX,
  47     TREG_EDX,
  48     TREG_EBX,
  49     TREG_ST0,
  50     TREG_ESP = 4
  51 };
  52
  53 /* return registers for function */
  54 #define REG_IRET TREG_EAX /* single word int return register */
  55 #define REG_IRE2 TREG_EDX /* second word return register (for long long) */
  56 #define REG_FRET TREG_ST0 /* float return register */
  57
  58 /* defined if function parameters must be evaluated in reverse order */
  59 #define INVERT_FUNC_PARAMS
  60
  61 /* defined if structures are passed as pointers. Otherwise structures
  62    are directly pushed on stack. */
  63 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  64
  65 /* pointer size, in bytes */
  66 #define PTR_SIZE 4
  67
  68 /* long double size and alignment, in bytes */
  69 #define LDOUBLE_SIZE  12
  70 #define LDOUBLE_ALIGN 4
  71 /* maximum alignment (for aligned attribute support) */
  72 #define MAX_ALIGN     8
  73
  74 /******************************************************/
  75 #else /* ! TARGET_DEFS_ONLY */
  76 /******************************************************/
  77 #define USING_GLOBALS
  78 #include "tcc.h"
  79
  80 /* define to 1/0 to [not] have EBX as 4th register */
  81 #define USE_EBX 0
  82
  83 ST_DATA const int reg_classes[NB_REGS] = {
  84     /* eax */ RC_INT | RC_EAX,
  85     /* ecx */ RC_INT | RC_ECX,
  86     /* edx */ RC_INT | RC_EDX,
  87     /* ebx */ (RC_INT | RC_EBX) * USE_EBX,
  88     /* st0 */ RC_FLOAT | RC_ST0,
  89 };
  90
  91 static unsigned long func_sub_sp_offset;
  92 static int func_ret_sub;
  93 #ifdef CONFIG_TCC_BCHECK
  94 static addr_t func_bound_offset;
  95 static unsigned long func_bound_ind;
  96 #endif
  97
  98 /* XXX: make it faster ? */
  99 ST_FUNC void g(int c)
 100 {
 101     int ind1;
 102     if (nocode_wanted)
 103         return;
 104     ind1 = ind + 1;
 105     if (ind1 > cur_text_section->data_allocated)
 106         section_realloc(cur_text_section, ind1);
 107     cur_text_section->data[ind] = c;
 108     ind = ind1;
 109 }
 110
 111 ST_FUNC void o(unsigned int c)
 112 {
 113     while (c) {
 114         g(c);
 115         c = c >> 8;
 116     }
 117 }
 118
 119 ST_FUNC void gen_le16(int v)
 120 {
 121     g(v);
 122     g(v >> 8);
 123 }
 124
 125 ST_FUNC void gen_le32(int c)
 126 {
 127     g(c);
 128     g(c >> 8);
 129     g(c >> 16);
 130     g(c >> 24);
 131 }
 132
 133 /* output a symbol and patch all calls to it */
 134 ST_FUNC void gsym_addr(int t, int a)
 135 {
 136     while (t) {
 137         unsigned char *ptr = cur_text_section->data + t;
 138         uint32_t n = read32le(ptr); /* next value */
 139         write32le(ptr, a - t - 4);
 140         t = n;
 141     }
 142 }
 143
 144 /* instruction + 4 bytes data. Return the address of the data */
 145 static int oad(int c, int s)
 146 {
 147     int t;
 148     if (nocode_wanted)
 149         return s;
 150     o(c);
 151     t = ind;
 152     gen_le32(s);
 153     return t;
 154 }
 155
 156 ST_FUNC void gen_fill_nops(int bytes)
 157 {
 158     while (bytes--)
 159       g(0x90);
 160 }
 161
 162 /* generate jmp to a label */
 163 #define gjmp2(instr,lbl) oad(instr,lbl)
 164
 165 /* output constant with relocation if 'r & VT_SYM' is true */
 166 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 167 {
 168     if (r & VT_SYM)
 169         greloc(cur_text_section, sym, ind, R_386_32);
 170     gen_le32(c);
 171 }
 172
 173 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 174 {
 175     if (r & VT_SYM)
 176         greloc(cur_text_section, sym, ind, R_386_PC32);
 177     gen_le32(c - 4);
 178 }
 179
 180 /* generate a modrm reference. 'op_reg' contains the additional 3
 181    opcode bits */
 182 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 183 {
 184     op_reg = op_reg << 3;
 185     if ((r & VT_VALMASK) == VT_CONST) {
 186         /* constant memory reference */
 187         o(0x05 | op_reg);
 188         gen_addr32(r, sym, c);
 189     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 190         /* currently, we use only ebp as base */
 191         if (c == (char)c) {
 192             /* short reference */
 193             o(0x45 | op_reg);
 194             g(c);
 195         } else {
 196             oad(0x85 | op_reg, c);
 197         }
 198     } else {
 199         g(0x00 | op_reg | (r & VT_VALMASK));
 200     }
 201 }
 202
 203 /* load 'r' from value 'sv' */
 204 ST_FUNC void load(int r, SValue *sv)
 205 {
 206     int v, t, ft, fc, fr;
 207     SValue v1;
 208
 209 #ifdef TCC_TARGET_PE
 210     SValue v2;
 211     sv = pe_getimport(sv, &v2);
 212 #endif
 213
 214     fr = sv->r;
 215     ft = sv->type.t & ~VT_DEFSIGN;
 216     fc = sv->c.i;
 217
 218     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 219
 220     v = fr & VT_VALMASK;
 221     if (fr & VT_LVAL) {
 222         if (v == VT_LLOCAL) {
 223             v1.type.t = VT_INT;
 224             v1.r = VT_LOCAL | VT_LVAL;
 225             v1.c.i = fc;
 226             fr = r;
 227             if (!(reg_classes[fr] & RC_INT))
 228                 fr = get_reg(RC_INT);
 229             load(fr, &v1);
 230         }
 231         if ((ft & VT_BTYPE) == VT_FLOAT) {
 232             o(0xd9); /* flds */
 233             r = 0;
 234         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 235             o(0xdd); /* fldl */
 236             r = 0;
 237         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 238             o(0xdb); /* fldt */
 239             r = 5;
 240         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 241             o(0xbe0f);   /* movsbl */
 242         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 243             o(0xb60f);   /* movzbl */
 244         } else if ((ft & VT_TYPE) == VT_SHORT) {
 245             o(0xbf0f);   /* movswl */
 246         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 247             o(0xb70f);   /* movzwl */
 248         } else {
 249             o(0x8b);     /* movl */
 250         }
 251         gen_modrm(r, fr, sv->sym, fc);
 252     } else {
 253         if (v == VT_CONST) {
 254             o(0xb8 + r); /* mov $xx, r */
 255             gen_addr32(fr, sv->sym, fc);
 256         } else if (v == VT_LOCAL) {
 257             if (fc) {
 258                 o(0x8d); /* lea xxx(%ebp), r */
 259                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 260             } else {
 261                 o(0x89);
 262                 o(0xe8 + r); /* mov %ebp, r */
 263             }
 264         } else if (v == VT_CMP) {
 265             o(0x0f); /* setxx %br */
 266             o(fc);
 267             o(0xc0 + r);
 268             o(0xc0b60f + r * 0x90000); /* movzbl %al, %eax */
 269         } else if (v == VT_JMP || v == VT_JMPI) {
 270             t = v & 1;
 271             oad(0xb8 + r, t); /* mov $1, r */
 272             o(0x05eb); /* jmp after */
 273             gsym(fc);
 274             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 275         } else if (v != r) {
 276             o(0x89);
 277             o(0xc0 + r + v * 8); /* mov v, r */
 278         }
 279     }
 280 }
 281
 282 /* store register 'r' in lvalue 'v' */
 283 ST_FUNC void store(int r, SValue *v)
 284 {
 285     int fr, bt, ft, fc;
 286
 287 #ifdef TCC_TARGET_PE
 288     SValue v2;
 289     v = pe_getimport(v, &v2);
 290 #endif
 291
 292     ft = v->type.t;
 293     fc = v->c.i;
 294     fr = v->r & VT_VALMASK;
 295     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 296     bt = ft & VT_BTYPE;
 297     /* XXX: incorrect if float reg to reg */
 298     if (bt == VT_FLOAT) {
 299         o(0xd9); /* fsts */
 300         r = 2;
 301     } else if (bt == VT_DOUBLE) {
 302         o(0xdd); /* fstpl */
 303         r = 2;
 304     } else if (bt == VT_LDOUBLE) {
 305         o(0xc0d9); /* fld %st(0) */
 306         o(0xdb); /* fstpt */
 307         r = 7;
 308     } else {
 309         if (bt == VT_SHORT)
 310             o(0x66);
 311         if (bt == VT_BYTE || bt == VT_BOOL)
 312             o(0x88);
 313         else
 314             o(0x89);
 315     }
 316     if (fr == VT_CONST ||
 317         fr == VT_LOCAL ||
 318         (v->r & VT_LVAL)) {
 319         gen_modrm(r, v->r, v->sym, fc);
 320     } else if (fr != r) {
 321         o(0xc0 + fr + r * 8); /* mov r, fr */
 322     }
 323 }
 324
 325 static void gadd_sp(int val)
 326 {
 327     if (val == (char)val) {
 328         o(0xc483);
 329         g(val);
 330     } else {
 331         oad(0xc481, val); /* add $xxx, %esp */
 332     }
 333 }
 334
 335 #if defined CONFIG_TCC_BCHECK || defined TCC_TARGET_PE
 336 static void gen_static_call(int v)
 337 {
 338     Sym *sym;
 339
 340     sym = external_global_sym(v, &func_old_type);
 341     oad(0xe8, -4);
 342     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 343 }
 344 #endif
 345
 346 /* 'is_jmp' is '1' if it is a jump */
 347 static void gcall_or_jmp(int is_jmp)
 348 {
 349     int r;
 350     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST && (vtop->r & VT_SYM)) {
 351         /* constant and relocation case */
 352         greloc(cur_text_section, vtop->sym, ind + 1, R_386_PC32);
 353         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 354     } else {
 355         /* otherwise, indirect call */
 356         r = gv(RC_INT);
 357         o(0xff); /* call/jmp *r */
 358         o(0xd0 + r + (is_jmp << 4));
 359     }
 360     if (!is_jmp) {
 361         int rt;
 362         /* extend the return value to the whole register if necessary
 363            visual studio and gcc do not always set the whole eax register
 364            when assigning the return value of a function  */
 365         rt = vtop->type.ref->type.t;
 366         switch (rt & VT_BTYPE) {
 367             case VT_BYTE:
 368             case VT_BOOL:
 369                 if (rt & VT_UNSIGNED) {
 370                     o(0xc0b60f); /* movzx %al, %eax */
 371                 }
 372                 else {
 373                     o(0xc0be0f); /* movsx %al, %eax */
 374                 }
 375                 break;
 376             case VT_SHORT:
 377                 if (rt & VT_UNSIGNED) {
 378                     o(0xc0b70f); /* movzx %ax, %eax */
 379                 }
 380                 else {
 381                     o(0xc0bf0f); /* movsx %ax, %eax */
 382                 }
 383                 break;
 384             default:
 385                 break;
 386         }
 387     }
 388 }
 389
 390 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 391 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 392
 393 /* Return the number of registers needed to return the struct, or 0 if
 394    returning via struct pointer. */
 395 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 396 {
 397 #ifdef TCC_TARGET_PE
 398     int size, align;
 399     *ret_align = 1; // Never have to re-align return values for x86
 400     *regsize = 4;
 401     size = type_size(vt, &align);
 402     if (size > 8 || (size & (size - 1)))
 403         return 0;
 404     if (size == 8)
 405         ret->t = VT_LLONG;
 406     else if (size == 4)
 407         ret->t = VT_INT;
 408     else if (size == 2)
 409         ret->t = VT_SHORT;
 410     else
 411         ret->t = VT_BYTE;
 412     ret->ref = NULL;
 413     return 1;
 414 #else
 415     *ret_align = 1; // Never have to re-align return values for x86
 416     return 0;
 417 #endif
 418 }
 419
 420 /* Generate function call. The function address is pushed first, then
 421    all the parameters in call order. This functions pops all the
 422    parameters and the function address. */
 423 ST_FUNC void gfunc_call(int nb_args)
 424 {
 425     int size, align, r, args_size, i, func_call;
 426     Sym *func_sym;
 427
 428 #ifdef CONFIG_TCC_BCHECK
 429     if (tcc_state->do_bounds_check)
 430         gbound_args(nb_args);
 431 #endif
 432
 433     args_size = 0;
 434     for(i = 0;i < nb_args; i++) {
 435         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 436             size = type_size(&vtop->type, &align);
 437             /* align to stack align size */
 438             size = (size + 3) & ~3;
 439             /* allocate the necessary size on stack */
 440             oad(0xec81, size); /* sub $xxx, %esp */
 441             /* generate structure store */
 442             r = get_reg(RC_INT);
 443             o(0x89); /* mov %esp, r */
 444             o(0xe0 + r);
 445             vset(&vtop->type, r | VT_LVAL, 0);
 446             vswap();
 447             vstore();
 448             args_size += size;
 449         } else if (is_float(vtop->type.t)) {
 450             gv(RC_FLOAT); /* only one float register */
 451             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 452                 size = 4;
 453             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 454                 size = 8;
 455             else
 456                 size = 12;
 457             oad(0xec81, size); /* sub $xxx, %esp */
 458             if (size == 12)
 459                 o(0x7cdb);
 460             else
 461                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 462             g(0x24);
 463             g(0x00);
 464             args_size += size;
 465         } else {
 466             /* simple type (currently always same size) */
 467             /* XXX: implicit cast ? */
 468             r = gv(RC_INT);
 469             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 470                 size = 8;
 471                 o(0x50 + vtop->r2); /* push r */
 472             } else {
 473                 size = 4;
 474             }
 475             o(0x50 + r); /* push r */
 476             args_size += size;
 477         }
 478         vtop--;
 479     }
 480     save_regs(0); /* save used temporary registers */
 481     func_sym = vtop->type.ref;
 482     func_call = func_sym->f.func_call;
 483     /* fast call case */
 484     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 485         func_call == FUNC_FASTCALLW) {
 486         int fastcall_nb_regs;
 487         uint8_t *fastcall_regs_ptr;
 488         if (func_call == FUNC_FASTCALLW) {
 489             fastcall_regs_ptr = fastcallw_regs;
 490             fastcall_nb_regs = 2;
 491         } else {
 492             fastcall_regs_ptr = fastcall_regs;
 493             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 494         }
 495         for(i = 0;i < fastcall_nb_regs; i++) {
 496             if (args_size <= 0)
 497                 break;
 498             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 499             /* XXX: incorrect for struct/floats */
 500             args_size -= 4;
 501         }
 502     }
 503 #ifndef TCC_TARGET_PE
 504     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 505         args_size -= 4;
 506 #endif
 507     gcall_or_jmp(0);
 508
 509     if (args_size && func_call != FUNC_STDCALL && func_call != FUNC_FASTCALLW)
 510         gadd_sp(args_size);
 511     vtop--;
 512 }
 513
 514 #ifdef TCC_TARGET_PE
 515 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 516 #else
 517 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 518 #endif
 519
 520 /* generate function prolog of type 't' */
 521 ST_FUNC void gfunc_prolog(Sym *func_sym)
 522 {
 523     CType *func_type = &func_sym->type;
 524     int addr, align, size, func_call, fastcall_nb_regs;
 525     int param_index, param_addr;
 526     int n_arg = 0;
 527     uint8_t *fastcall_regs_ptr;
 528     Sym *sym;
 529     CType *type;
 530
 531     sym = func_type->ref;
 532     func_call = sym->f.func_call;
 533     addr = 8;
 534     loc = 0;
 535     func_vc = 0;
 536
 537     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 538         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 539         fastcall_regs_ptr = fastcall_regs;
 540     } else if (func_call == FUNC_FASTCALLW) {
 541         fastcall_nb_regs = 2;
 542         fastcall_regs_ptr = fastcallw_regs;
 543     } else {
 544         fastcall_nb_regs = 0;
 545         fastcall_regs_ptr = NULL;
 546     }
 547     param_index = 0;
 548
 549     ind += FUNC_PROLOG_SIZE;
 550     func_sub_sp_offset = ind;
 551     /* if the function returns a structure, then add an
 552        implicit pointer parameter */
 553     func_vt = sym->type;
 554     func_var = (sym->f.func_type == FUNC_ELLIPSIS);
 555 #ifdef TCC_TARGET_PE
 556     size = type_size(&func_vt,&align);
 557     if (((func_vt.t & VT_BTYPE) == VT_STRUCT)
 558         && (size > 8 || (size & (size - 1)))) {
 559 #else
 560     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 561 #endif
 562         /* XXX: fastcall case ? */
 563         func_vc = addr;
 564         addr += 4;
 565         param_index++;
 566     }
 567     /* define parameters */
 568     while ((sym = sym->next) != NULL) {
 569         n_arg++;
 570         type = &sym->type;
 571         size = type_size(type, &align);
 572         size = (size + 3) & ~3;
 573 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 574         /* structs are passed as pointer */
 575         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 576             size = 4;
 577         }
 578 #endif
 579         if (param_index < fastcall_nb_regs) {
 580             /* save FASTCALL register */
 581             loc -= 4;
 582             o(0x89);     /* movl */
 583             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 584             param_addr = loc;
 585         } else {
 586             param_addr = addr;
 587             addr += size;
 588         }
 589         sym_push(sym->v & ~SYM_FIELD, type,
 590                  VT_LOCAL | VT_LVAL, param_addr);
 591         param_index++;
 592     }
 593     func_ret_sub = 0;
 594     /* pascal type call or fastcall ? */
 595     if (func_call == FUNC_STDCALL || func_call == FUNC_FASTCALLW)
 596         func_ret_sub = addr - 8;
 597 #ifndef TCC_TARGET_PE
 598     else if (func_vc)
 599         func_ret_sub = 4;
 600 #endif
 601
 602 #ifdef CONFIG_TCC_BCHECK
 603     /* leave some room for bound checking code */
 604     if (tcc_state->do_bounds_check) {
 605         func_bound_offset = lbounds_section->data_offset;
 606         func_bound_ind = ind;
 607         oad(0xb8, 0); /* lbound section pointer */
 608         oad(0xb8, 0); /* call to function */
 609         if (n_arg >= 2 && strcmp (get_tok_str(func_sym->v, NULL), "main") == 0) {
 610             o(0x0c458b);  /* mov  0x12(%ebp),%eax */
 611             o(0x50);      /* push %eax */
 612             gen_static_call(TOK___bound_main_arg);
 613             o(0x04c483);  /* add  $0x4,%esp */
 614         }
 615     }
 616 #endif
 617 }
 618
 619 /* generate function epilog */
 620 ST_FUNC void gfunc_epilog(void)
 621 {
 622     addr_t v, saved_ind;
 623
 624 #ifdef CONFIG_TCC_BCHECK
 625     if (tcc_state->do_bounds_check
 626      && func_bound_offset != lbounds_section->data_offset) {
 627         addr_t saved_ind;
 628         addr_t *bounds_ptr;
 629         Sym *sym_data;
 630
 631         /* add end of table info */
 632         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 633         *bounds_ptr = 0;
 634
 635         /* generate bound local allocation */
 636         saved_ind = ind;
 637         ind = func_bound_ind;
 638         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 639                                func_bound_offset, lbounds_section->data_offset);
 640         greloc(cur_text_section, sym_data,
 641                ind + 1, R_386_32);
 642         oad(0xb8, 0); /* mov %eax, xxx */
 643         gen_static_call(TOK___bound_local_new);
 644         ind = saved_ind;
 645
 646         /* generate bound check local freeing */
 647         o(0x5250); /* save returned value, if any */
 648         greloc(cur_text_section, sym_data, ind + 1, R_386_32);
 649         oad(0xb8, 0); /* mov %eax, xxx */
 650         gen_static_call(TOK___bound_local_delete);
 651         o(0x585a); /* restore returned value, if any */
 652     }
 653 #endif
 654
 655     /* align local size to word & save local variables */
 656     v = (-loc + 3) & -4;
 657
 658 #if USE_EBX
 659     o(0x8b);
 660     gen_modrm(TREG_EBX, VT_LOCAL, NULL, -(v+4));
 661 #endif
 662
 663     o(0xc9); /* leave */
 664     if (func_ret_sub == 0) {
 665         o(0xc3); /* ret */
 666     } else {
 667         o(0xc2); /* ret n */
 668         g(func_ret_sub);
 669         g(func_ret_sub >> 8);
 670     }
 671     saved_ind = ind;
 672     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 673 #ifdef TCC_TARGET_PE
 674     if (v >= 4096) {
 675         oad(0xb8, v); /* mov stacksize, %eax */
 676         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 677     } else
 678 #endif
 679     {
 680         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 681         o(0xec81);  /* sub esp, stacksize */
 682         gen_le32(v);
 683 #ifdef TCC_TARGET_PE
 684         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 685 #endif
 686     }
 687     o(0x53 * USE_EBX); /* push ebx */
 688     ind = saved_ind;
 689 }
 690
 691 /* generate a jump to a label */
 692 ST_FUNC int gjmp(int t)
 693 {
 694     return gjmp2(0xe9, t);
 695 }
 696
 697 /* generate a jump to a fixed address */
 698 ST_FUNC void gjmp_addr(int a)
 699 {
 700     int r;
 701     r = a - ind - 2;
 702     if (r == (char)r) {
 703         g(0xeb);
 704         g(r);
 705     } else {
 706         oad(0xe9, a - ind - 5);
 707     }
 708 }
 709
 710 #if 0
 711 /* generate a jump to a fixed address */
 712 ST_FUNC void gjmp_cond_addr(int a, int op)
 713 {
 714     int r = a - ind - 2;
 715     if (r == (char)r)
 716         g(op - 32), g(r);
 717     else
 718         g(0x0f), gjmp2(op - 16, r - 4);
 719 }
 720 #endif
 721
 722 ST_FUNC int gjmp_append(int n, int t)
 723 {
 724     void *p;
 725     /* insert vtop->c jump list in t */
 726     if (n) {
 727         uint32_t n1 = n, n2;
 728         while ((n2 = read32le(p = cur_text_section->data + n1)))
 729             n1 = n2;
 730         write32le(p, t);
 731         t = n;
 732     }
 733     return t;
 734 }
 735
 736 ST_FUNC int gjmp_cond(int op, int t)
 737 {
 738     g(0x0f);
 739     t = gjmp2(op - 16, t);
 740     return t;
 741 }
 742
 743 ST_FUNC void gen_opi(int op)
 744 {
 745     int r, fr, opc, c;
 746
 747     switch(op) {
 748     case '+':
 749     case TOK_ADDC1: /* add with carry generation */
 750         opc = 0;
 751     gen_op8:
 752         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 753             /* constant case */
 754             vswap();
 755             r = gv(RC_INT);
 756             vswap();
 757             c = vtop->c.i;
 758             if (c == (char)c) {
 759                 /* generate inc and dec for smaller code */
 760                 if ((c == 1 || c == -1) && (op == '+' || op == '-')) {
 761                     opc = (c == 1) ^ (op == '+');
 762                     o (0x40 | (opc << 3) | r); // inc,dec
 763                 } else {
 764                     o(0x83);
 765                     o(0xc0 | (opc << 3) | r);
 766                     g(c);
 767                 }
 768             } else {
 769                 o(0x81);
 770                 oad(0xc0 | (opc << 3) | r, c);
 771             }
 772         } else {
 773             gv2(RC_INT, RC_INT);
 774             r = vtop[-1].r;
 775             fr = vtop[0].r;
 776             o((opc << 3) | 0x01);
 777             o(0xc0 + r + fr * 8);
 778         }
 779         vtop--;
 780         if (op >= TOK_ULT && op <= TOK_GT)
 781             vset_VT_CMP(op);
 782         break;
 783     case '-':
 784     case TOK_SUBC1: /* sub with carry generation */
 785         opc = 5;
 786         goto gen_op8;
 787     case TOK_ADDC2: /* add with carry use */
 788         opc = 2;
 789         goto gen_op8;
 790     case TOK_SUBC2: /* sub with carry use */
 791         opc = 3;
 792         goto gen_op8;
 793     case '&':
 794         opc = 4;
 795         goto gen_op8;
 796     case '^':
 797         opc = 6;
 798         goto gen_op8;
 799     case '|':
 800         opc = 1;
 801         goto gen_op8;
 802     case '*':
 803         gv2(RC_INT, RC_INT);
 804         r = vtop[-1].r;
 805         fr = vtop[0].r;
 806         vtop--;
 807         o(0xaf0f); /* imul fr, r */
 808         o(0xc0 + fr + r * 8);
 809         break;
 810     case TOK_SHL:
 811         opc = 4;
 812         goto gen_shift;
 813     case TOK_SHR:
 814         opc = 5;
 815         goto gen_shift;
 816     case TOK_SAR:
 817         opc = 7;
 818     gen_shift:
 819         opc = 0xc0 | (opc << 3);
 820         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 821             /* constant case */
 822             vswap();
 823             r = gv(RC_INT);
 824             vswap();
 825             c = vtop->c.i & 0x1f;
 826             o(0xc1); /* shl/shr/sar $xxx, r */
 827             o(opc | r);
 828             g(c);
 829         } else {
 830             /* we generate the shift in ecx */
 831             gv2(RC_INT, RC_ECX);
 832             r = vtop[-1].r;
 833             o(0xd3); /* shl/shr/sar %cl, r */
 834             o(opc | r);
 835         }
 836         vtop--;
 837         break;
 838     case '/':
 839     case TOK_UDIV:
 840     case TOK_PDIV:
 841     case '%':
 842     case TOK_UMOD:
 843     case TOK_UMULL:
 844         /* first operand must be in eax */
 845         /* XXX: need better constraint for second operand */
 846         gv2(RC_EAX, RC_ECX);
 847         r = vtop[-1].r;
 848         fr = vtop[0].r;
 849         vtop--;
 850         save_reg(TREG_EDX);
 851         /* save EAX too if used otherwise */
 852         save_reg_upstack(TREG_EAX, 1);
 853         if (op == TOK_UMULL) {
 854             o(0xf7); /* mul fr */
 855             o(0xe0 + fr);
 856             vtop->r2 = TREG_EDX;
 857             r = TREG_EAX;
 858         } else {
 859             if (op == TOK_UDIV || op == TOK_UMOD) {
 860                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 861                 o(0xf0 + fr);
 862             } else {
 863                 o(0xf799); /* cltd, idiv fr, %eax */
 864                 o(0xf8 + fr);
 865             }
 866             if (op == '%' || op == TOK_UMOD)
 867                 r = TREG_EDX;
 868             else
 869                 r = TREG_EAX;
 870         }
 871         vtop->r = r;
 872         break;
 873     default:
 874         opc = 7;
 875         goto gen_op8;
 876     }
 877 }
 878
 879 /* generate a floating point operation 'v = t1 op t2' instruction. The
 880    two operands are guaranteed to have the same floating point type */
 881 /* XXX: need to use ST1 too */
 882 ST_FUNC void gen_opf(int op)
 883 {
 884     int a, ft, fc, swapped, r;
 885
 886     /* convert constants to memory references */
 887     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 888         vswap();
 889         gv(RC_FLOAT);
 890         vswap();
 891     }
 892     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 893         gv(RC_FLOAT);
 894
 895     /* must put at least one value in the floating point register */
 896     if ((vtop[-1].r & VT_LVAL) &&
 897         (vtop[0].r & VT_LVAL)) {
 898         vswap();
 899         gv(RC_FLOAT);
 900         vswap();
 901     }
 902     swapped = 0;
 903     /* swap the stack if needed so that t1 is the register and t2 is
 904        the memory reference */
 905     if (vtop[-1].r & VT_LVAL) {
 906         vswap();
 907         swapped = 1;
 908     }
 909     if (op >= TOK_ULT && op <= TOK_GT) {
 910         /* load on stack second operand */
 911         load(TREG_ST0, vtop);
 912         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 913         if (op == TOK_GE || op == TOK_GT)
 914             swapped = !swapped;
 915         else if (op == TOK_EQ || op == TOK_NE)
 916             swapped = 0;
 917         if (swapped)
 918             o(0xc9d9); /* fxch %st(1) */
 919         if (op == TOK_EQ || op == TOK_NE)
 920             o(0xe9da); /* fucompp */
 921         else
 922             o(0xd9de); /* fcompp */
 923         o(0xe0df); /* fnstsw %ax */
 924         if (op == TOK_EQ) {
 925             o(0x45e480); /* and $0x45, %ah */
 926             o(0x40fC80); /* cmp $0x40, %ah */
 927         } else if (op == TOK_NE) {
 928             o(0x45e480); /* and $0x45, %ah */
 929             o(0x40f480); /* xor $0x40, %ah */
 930             op = TOK_NE;
 931         } else if (op == TOK_GE || op == TOK_LE) {
 932             o(0x05c4f6); /* test $0x05, %ah */
 933             op = TOK_EQ;
 934         } else {
 935             o(0x45c4f6); /* test $0x45, %ah */
 936             op = TOK_EQ;
 937         }
 938         vtop--;
 939         vset_VT_CMP(op);
 940     } else {
 941         /* no memory reference possible for long double operations */
 942         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 943             load(TREG_ST0, vtop);
 944             swapped = !swapped;
 945         }
 946
 947         switch(op) {
 948         default:
 949         case '+':
 950             a = 0;
 951             break;
 952         case '-':
 953             a = 4;
 954             if (swapped)
 955                 a++;
 956             break;
 957         case '*':
 958             a = 1;
 959             break;
 960         case '/':
 961             a = 6;
 962             if (swapped)
 963                 a++;
 964             break;
 965         }
 966         ft = vtop->type.t;
 967         fc = vtop->c.i;
 968         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 969             o(0xde); /* fxxxp %st, %st(1) */
 970             o(0xc1 + (a << 3));
 971         } else {
 972             /* if saved lvalue, then we must reload it */
 973             r = vtop->r;
 974             if ((r & VT_VALMASK) == VT_LLOCAL) {
 975                 SValue v1;
 976                 r = get_reg(RC_INT);
 977                 v1.type.t = VT_INT;
 978                 v1.r = VT_LOCAL | VT_LVAL;
 979                 v1.c.i = fc;
 980                 load(r, &v1);
 981                 fc = 0;
 982             }
 983
 984             if ((ft & VT_BTYPE) == VT_DOUBLE)
 985                 o(0xdc);
 986             else
 987                 o(0xd8);
 988             gen_modrm(a, r, vtop->sym, fc);
 989         }
 990         vtop--;
 991     }
 992 }
 993
 994 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 995    and 'long long' cases. */
 996 ST_FUNC void gen_cvt_itof(int t)
 997 {
 998     save_reg(TREG_ST0);
 999     gv(RC_INT);
1000     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
1001         /* signed long long to float/double/long double (unsigned case
1002            is handled generically) */
1003         o(0x50 + vtop->r2); /* push r2 */
1004         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1005         o(0x242cdf); /* fildll (%esp) */
1006         o(0x08c483); /* add $8, %esp */
1007         vtop->r2 = VT_CONST;
1008     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
1009                (VT_INT | VT_UNSIGNED)) {
1010         /* unsigned int to float/double/long double */
1011         o(0x6a); /* push $0 */
1012         g(0x00);
1013         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1014         o(0x242cdf); /* fildll (%esp) */
1015         o(0x08c483); /* add $8, %esp */
1016     } else {
1017         /* int to float/double/long double */
1018         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1019         o(0x2404db); /* fildl (%esp) */
1020         o(0x04c483); /* add $4, %esp */
1021     }
1022     vtop->r2 = VT_CONST;
1023     vtop->r = TREG_ST0;
1024 }
1025
1026 /* convert fp to int 't' type */
1027 ST_FUNC void gen_cvt_ftoi(int t)
1028 {
1029     int bt = vtop->type.t & VT_BTYPE;
1030     if (bt == VT_FLOAT)
1031         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1032     else if (bt == VT_LDOUBLE)
1033         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1034     else
1035         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1036     vswap();
1037     gfunc_call(1);
1038     vpushi(0);
1039     vtop->r = REG_IRET;
1040     if ((t & VT_BTYPE) == VT_LLONG)
1041         vtop->r2 = REG_IRE2;
1042 }
1043
1044 /* convert from one floating point type to another */
1045 ST_FUNC void gen_cvt_ftof(int t)
1046 {
1047     /* all we have to do on i386 is to put the float in a register */
1048     gv(RC_FLOAT);
1049 }
1050
1051 /* computed goto support */
1052 ST_FUNC void ggoto(void)
1053 {
1054     gcall_or_jmp(1);
1055     vtop--;
1056 }
1057
1058 /* bound check support functions */
1059 #ifdef CONFIG_TCC_BCHECK
1060 /* generate a bounded pointer addition */
1061 ST_FUNC void gen_bounded_ptr_add(void)
1062 {
1063     vpush_global_sym(&func_old_type, TOK___bound_ptr_add);
1064     vrott(3);
1065     gfunc_call(2);
1066     vpushi(0);
1067     /* returned pointer is in eax */
1068     vtop->r = TREG_EAX | VT_BOUNDED;
1069     if (nocode_wanted)
1070         return;
1071     /* relocation offset of the bounding function call point */
1072     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1073 }
1074
1075 /* patch pointer addition in vtop so that pointer dereferencing is
1076    also tested */
1077 ST_FUNC void gen_bounded_ptr_deref(void)
1078 {
1079     addr_t func;
1080     int  size, align;
1081     Elf32_Rel *rel;
1082     Sym *sym;
1083
1084     if (nocode_wanted)
1085         return;
1086
1087     size = type_size(&vtop->type, &align);
1088     switch(size) {
1089     case  1: func = TOK___bound_ptr_indir1; break;
1090     case  2: func = TOK___bound_ptr_indir2; break;
1091     case  4: func = TOK___bound_ptr_indir4; break;
1092     case  8: func = TOK___bound_ptr_indir8; break;
1093     case 12: func = TOK___bound_ptr_indir12; break;
1094     case 16: func = TOK___bound_ptr_indir16; break;
1095     default:
1096         /* may happen with struct member access */
1097         return;
1098         //tcc_error("unhandled size when dereferencing bounded pointer");
1099         //func = 0;
1100         //break;
1101     }
1102     sym = external_global_sym(func, &func_old_type);
1103     if (!sym->c)
1104         put_extern_sym(sym, NULL, 0, 0);
1105     /* patch relocation */
1106     /* XXX: find a better solution ? */
1107     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1108     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1109 }
1110 #endif
1111
1112 /* Save the stack pointer onto the stack */
1113 ST_FUNC void gen_vla_sp_save(int addr) {
1114     /* mov %esp,addr(%ebp)*/
1115     o(0x89);
1116     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1117 }
1118
1119 /* Restore the SP from a location on the stack */
1120 ST_FUNC void gen_vla_sp_restore(int addr) {
1121     o(0x8b);
1122     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1123 }
1124
1125 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1126 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1127     int use_call = 0;
1128
1129 #if defined(CONFIG_TCC_BCHECK)
1130     use_call = tcc_state->do_bounds_check;
1131 #endif
1132 #ifdef TCC_TARGET_PE    /* alloca does more than just adjust %rsp on Windows */
1133     use_call = 1;
1134 #endif
1135     if (use_call)
1136     {
1137         vpush_global_sym(&func_old_type, TOK_alloca);
1138         vswap(); /* Move alloca ref past allocation size */
1139         gfunc_call(1);
1140     }
1141     else {
1142         int r;
1143         r = gv(RC_INT); /* allocation size */
1144         /* sub r,%rsp */
1145         o(0x2b);
1146         o(0xe0 | r);
1147         /* We align to 16 bytes rather than align */
1148         /* and ~15, %esp */
1149         o(0xf0e483);
1150         vpop();
1151     }
1152 }
1153
1154 /* end of X86 code generator */
1155 /*************************************************************/
1156 #endif
1157 /*************************************************************/