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