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_LRET    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_LRET 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 #include "tcc.h"
  78
  79 /* define to 1/0 to [not] have EBX as 4th register */
  80 #define USE_EBX 0
  81
  82 ST_DATA const int reg_classes[NB_REGS] = {
  83     /* eax */ RC_INT | RC_EAX,
  84     /* ecx */ RC_INT | RC_ECX,
  85     /* edx */ RC_INT | RC_EDX,
  86     /* ebx */ (RC_INT | RC_EBX) * USE_EBX,
  87     /* st0 */ RC_FLOAT | RC_ST0,
  88 };
  89
  90 static unsigned long func_sub_sp_offset;
  91 static int func_ret_sub;
  92 #ifdef CONFIG_TCC_BCHECK
  93 static addr_t func_bound_offset;
  94 static unsigned long func_bound_ind;
  95 #endif
  96
  97 /* XXX: make it faster ? */
  98 ST_FUNC void g(int c)
  99 {
 100     int ind1;
 101     if (nocode_wanted)
 102         return;
 103     ind1 = ind + 1;
 104     if (ind1 > cur_text_section->data_allocated)
 105         section_realloc(cur_text_section, ind1);
 106     cur_text_section->data[ind] = c;
 107     ind = ind1;
 108 }
 109
 110 ST_FUNC void o(unsigned int c)
 111 {
 112     while (c) {
 113         g(c);
 114         c = c >> 8;
 115     }
 116 }
 117
 118 ST_FUNC void gen_le16(int v)
 119 {
 120     g(v);
 121     g(v >> 8);
 122 }
 123
 124 ST_FUNC void gen_le32(int c)
 125 {
 126     g(c);
 127     g(c >> 8);
 128     g(c >> 16);
 129     g(c >> 24);
 130 }
 131
 132 /* output a symbol and patch all calls to it */
 133 ST_FUNC void gsym_addr(int t, int a)
 134 {
 135     while (t) {
 136         unsigned char *ptr = cur_text_section->data + t;
 137         uint32_t n = read32le(ptr); /* next value */
 138         write32le(ptr, a - t - 4);
 139         t = n;
 140     }
 141 }
 142
 143 ST_FUNC void gsym(int t)
 144 {
 145     gsym_addr(t, ind);
 146 }
 147
 148 /* instruction + 4 bytes data. Return the address of the data */
 149 static int oad(int c, int s)
 150 {
 151     int t;
 152     if (nocode_wanted)
 153         return s;
 154     o(c);
 155     t = ind;
 156     gen_le32(s);
 157     return t;
 158 }
 159
 160 /* generate jmp to a label */
 161 #define gjmp2(instr,lbl) oad(instr,lbl)
 162
 163 /* output constant with relocation if 'r & VT_SYM' is true */
 164 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 165 {
 166     if (r & VT_SYM)
 167         greloc(cur_text_section, sym, ind, R_386_32);
 168     gen_le32(c);
 169 }
 170
 171 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 172 {
 173     if (r & VT_SYM)
 174         greloc(cur_text_section, sym, ind, R_386_PC32);
 175     gen_le32(c - 4);
 176 }
 177
 178 /* generate a modrm reference. 'op_reg' contains the additional 3
 179    opcode bits */
 180 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 181 {
 182     op_reg = op_reg << 3;
 183     if ((r & VT_VALMASK) == VT_CONST) {
 184         /* constant memory reference */
 185         o(0x05 | op_reg);
 186         gen_addr32(r, sym, c);
 187     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 188         /* currently, we use only ebp as base */
 189         if (c == (char)c) {
 190             /* short reference */
 191             o(0x45 | op_reg);
 192             g(c);
 193         } else {
 194             oad(0x85 | op_reg, c);
 195         }
 196     } else {
 197         g(0x00 | op_reg | (r & VT_VALMASK));
 198     }
 199 }
 200
 201 /* load 'r' from value 'sv' */
 202 ST_FUNC void load(int r, SValue *sv)
 203 {
 204     int v, t, ft, fc, fr;
 205     SValue v1;
 206
 207 #ifdef TCC_TARGET_PE
 208     SValue v2;
 209     sv = pe_getimport(sv, &v2);
 210 #endif
 211
 212     fr = sv->r;
 213     ft = sv->type.t & ~VT_DEFSIGN;
 214     fc = sv->c.i;
 215
 216     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 217
 218     v = fr & VT_VALMASK;
 219     if (fr & VT_LVAL) {
 220         if (v == VT_LLOCAL) {
 221             v1.type.t = VT_INT;
 222             v1.r = VT_LOCAL | VT_LVAL;
 223             v1.c.i = fc;
 224             fr = r;
 225             if (!(reg_classes[fr] & RC_INT))
 226                 fr = get_reg(RC_INT);
 227             load(fr, &v1);
 228         }
 229         if ((ft & VT_BTYPE) == VT_FLOAT) {
 230             o(0xd9); /* flds */
 231             r = 0;
 232         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 233             o(0xdd); /* fldl */
 234             r = 0;
 235         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 236             o(0xdb); /* fldt */
 237             r = 5;
 238         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 239             o(0xbe0f);   /* movsbl */
 240         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 241             o(0xb60f);   /* movzbl */
 242         } else if ((ft & VT_TYPE) == VT_SHORT) {
 243             o(0xbf0f);   /* movswl */
 244         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 245             o(0xb70f);   /* movzwl */
 246         } else {
 247             o(0x8b);     /* movl */
 248         }
 249         gen_modrm(r, fr, sv->sym, fc);
 250     } else {
 251         if (v == VT_CONST) {
 252             o(0xb8 + r); /* mov $xx, r */
 253             gen_addr32(fr, sv->sym, fc);
 254         } else if (v == VT_LOCAL) {
 255             if (fc) {
 256                 o(0x8d); /* lea xxx(%ebp), r */
 257                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 258             } else {
 259                 o(0x89);
 260                 o(0xe8 + r); /* mov %ebp, r */
 261             }
 262         } else if (v == VT_CMP) {
 263             oad(0xb8 + r, 0); /* mov $0, r */
 264             o(0x0f); /* setxx %br */
 265             o(fc);
 266             o(0xc0 + r);
 267         } else if (v == VT_JMP || v == VT_JMPI) {
 268             t = v & 1;
 269             oad(0xb8 + r, t); /* mov $1, r */
 270             o(0x05eb); /* jmp after */
 271             gsym(fc);
 272             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 273         } else if (v != r) {
 274             o(0x89);
 275             o(0xc0 + r + v * 8); /* mov v, r */
 276         }
 277     }
 278 }
 279
 280 /* store register 'r' in lvalue 'v' */
 281 ST_FUNC void store(int r, SValue *v)
 282 {
 283     int fr, bt, ft, fc;
 284
 285 #ifdef TCC_TARGET_PE
 286     SValue v2;
 287     v = pe_getimport(v, &v2);
 288 #endif
 289
 290     ft = v->type.t;
 291     fc = v->c.i;
 292     fr = v->r & VT_VALMASK;
 293     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 294     bt = ft & VT_BTYPE;
 295     /* XXX: incorrect if float reg to reg */
 296     if (bt == VT_FLOAT) {
 297         o(0xd9); /* fsts */
 298         r = 2;
 299     } else if (bt == VT_DOUBLE) {
 300         o(0xdd); /* fstpl */
 301         r = 2;
 302     } else if (bt == VT_LDOUBLE) {
 303         o(0xc0d9); /* fld %st(0) */
 304         o(0xdb); /* fstpt */
 305         r = 7;
 306     } else {
 307         if (bt == VT_SHORT)
 308             o(0x66);
 309         if (bt == VT_BYTE || bt == VT_BOOL)
 310             o(0x88);
 311         else
 312             o(0x89);
 313     }
 314     if (fr == VT_CONST ||
 315         fr == VT_LOCAL ||
 316         (v->r & VT_LVAL)) {
 317         gen_modrm(r, v->r, v->sym, fc);
 318     } else if (fr != r) {
 319         o(0xc0 + fr + r * 8); /* mov r, fr */
 320     }
 321 }
 322
 323 static void gadd_sp(int val)
 324 {
 325     if (val == (char)val) {
 326         o(0xc483);
 327         g(val);
 328     } else {
 329         oad(0xc481, val); /* add $xxx, %esp */
 330     }
 331 }
 332
 333 #if defined CONFIG_TCC_BCHECK || defined TCC_TARGET_PE
 334 static void gen_static_call(int v)
 335 {
 336     Sym *sym;
 337
 338     sym = external_global_sym(v, &func_old_type, 0);
 339     oad(0xe8, -4);
 340     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 341 }
 342 #endif
 343
 344 /* 'is_jmp' is '1' if it is a jump */
 345 static void gcall_or_jmp(int is_jmp)
 346 {
 347     int r;
 348     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST && (vtop->r & VT_SYM)) {
 349         /* constant and relocation case */
 350         greloc(cur_text_section, vtop->sym, ind + 1, R_386_PC32);
 351         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 352     } else {
 353         /* otherwise, indirect call */
 354         r = gv(RC_INT);
 355         o(0xff); /* call/jmp *r */
 356         o(0xd0 + r + (is_jmp << 4));
 357     }
 358     if (!is_jmp) {
 359         int rt;
 360         /* extend the return value to the whole register if necessary
 361            visual studio and gcc do not always set the whole eax register
 362            when assigning the return value of a function  */
 363         rt = vtop->type.ref->type.t;
 364         switch (rt & VT_BTYPE) {
 365             case VT_BYTE:
 366                 if (rt & VT_UNSIGNED) {
 367                     o(0xc0b60f); /* movzx %al, %eax */
 368                 }
 369                 else {
 370                     o(0xc0be0f); /* movsx %al, %eax */
 371                 }
 372                 break;
 373             case VT_SHORT:
 374                 if (rt & VT_UNSIGNED) {
 375                     o(0xc0b70f); /* movzx %ax, %eax */
 376                 }
 377                 else {
 378                     o(0xc0bf0f); /* movsx %ax, %eax */
 379                 }
 380                 break;
 381             default:
 382                 break;
 383         }
 384     }
 385 }
 386
 387 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 388 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 389
 390 /* Return the number of registers needed to return the struct, or 0 if
 391    returning via struct pointer. */
 392 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 393 {
 394 #ifdef TCC_TARGET_PE
 395     int size, align;
 396     *ret_align = 1; // Never have to re-align return values for x86
 397     *regsize = 4;
 398     size = type_size(vt, &align);
 399     if (size > 8 || (size & (size - 1)))
 400         return 0;
 401     if (size == 8)
 402         ret->t = VT_LLONG;
 403     else if (size == 4)
 404         ret->t = VT_INT;
 405     else if (size == 2)
 406         ret->t = VT_SHORT;
 407     else
 408         ret->t = VT_BYTE;
 409     ret->ref = NULL;
 410     return 1;
 411 #else
 412     *ret_align = 1; // Never have to re-align return values for x86
 413     return 0;
 414 #endif
 415 }
 416
 417 /* Generate function call. The function address is pushed first, then
 418    all the parameters in call order. This functions pops all the
 419    parameters and the function address. */
 420 ST_FUNC void gfunc_call(int nb_args)
 421 {
 422     int size, align, r, args_size, i, func_call;
 423     Sym *func_sym;
 424
 425     args_size = 0;
 426     for(i = 0;i < nb_args; i++) {
 427         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 428             size = type_size(&vtop->type, &align);
 429             /* align to stack align size */
 430             size = (size + 3) & ~3;
 431             /* allocate the necessary size on stack */
 432             oad(0xec81, size); /* sub $xxx, %esp */
 433             /* generate structure store */
 434             r = get_reg(RC_INT);
 435             o(0x89); /* mov %esp, r */
 436             o(0xe0 + r);
 437             vset(&vtop->type, r | VT_LVAL, 0);
 438             vswap();
 439             vstore();
 440             args_size += size;
 441         } else if (is_float(vtop->type.t)) {
 442             gv(RC_FLOAT); /* only one float register */
 443             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 444                 size = 4;
 445             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 446                 size = 8;
 447             else
 448                 size = 12;
 449             oad(0xec81, size); /* sub $xxx, %esp */
 450             if (size == 12)
 451                 o(0x7cdb);
 452             else
 453                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 454             g(0x24);
 455             g(0x00);
 456             args_size += size;
 457         } else {
 458             /* simple type (currently always same size) */
 459             /* XXX: implicit cast ? */
 460             r = gv(RC_INT);
 461             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 462                 size = 8;
 463                 o(0x50 + vtop->r2); /* push r */
 464             } else {
 465                 size = 4;
 466             }
 467             o(0x50 + r); /* push r */
 468             args_size += size;
 469         }
 470         vtop--;
 471     }
 472     save_regs(0); /* save used temporary registers */
 473     func_sym = vtop->type.ref;
 474     func_call = func_sym->f.func_call;
 475     /* fast call case */
 476     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 477         func_call == FUNC_FASTCALLW) {
 478         int fastcall_nb_regs;
 479         uint8_t *fastcall_regs_ptr;
 480         if (func_call == FUNC_FASTCALLW) {
 481             fastcall_regs_ptr = fastcallw_regs;
 482             fastcall_nb_regs = 2;
 483         } else {
 484             fastcall_regs_ptr = fastcall_regs;
 485             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 486         }
 487         for(i = 0;i < fastcall_nb_regs; i++) {
 488             if (args_size <= 0)
 489                 break;
 490             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 491             /* XXX: incorrect for struct/floats */
 492             args_size -= 4;
 493         }
 494     }
 495 #ifndef TCC_TARGET_PE
 496     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 497         args_size -= 4;
 498 #endif
 499     gcall_or_jmp(0);
 500
 501     if (args_size && func_call != FUNC_STDCALL && func_call != FUNC_FASTCALLW)
 502         gadd_sp(args_size);
 503     vtop--;
 504 }
 505
 506 #ifdef TCC_TARGET_PE
 507 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 508 #else
 509 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 510 #endif
 511
 512 /* generate function prolog of type 't' */
 513 ST_FUNC void gfunc_prolog(CType *func_type)
 514 {
 515     int addr, align, size, func_call, fastcall_nb_regs;
 516     int param_index, param_addr;
 517     uint8_t *fastcall_regs_ptr;
 518     Sym *sym;
 519     CType *type;
 520
 521     sym = func_type->ref;
 522     func_call = sym->f.func_call;
 523     addr = 8;
 524     loc = 0;
 525     func_vc = 0;
 526
 527     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 528         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 529         fastcall_regs_ptr = fastcall_regs;
 530     } else if (func_call == FUNC_FASTCALLW) {
 531         fastcall_nb_regs = 2;
 532         fastcall_regs_ptr = fastcallw_regs;
 533     } else {
 534         fastcall_nb_regs = 0;
 535         fastcall_regs_ptr = NULL;
 536     }
 537     param_index = 0;
 538
 539     ind += FUNC_PROLOG_SIZE;
 540     func_sub_sp_offset = ind;
 541     /* if the function returns a structure, then add an
 542        implicit pointer parameter */
 543     func_vt = sym->type;
 544     func_var = (sym->f.func_type == FUNC_ELLIPSIS);
 545 #ifdef TCC_TARGET_PE
 546     size = type_size(&func_vt,&align);
 547     if (((func_vt.t & VT_BTYPE) == VT_STRUCT)
 548         && (size > 8 || (size & (size - 1)))) {
 549 #else
 550     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 551 #endif
 552         /* XXX: fastcall case ? */
 553         func_vc = addr;
 554         addr += 4;
 555         param_index++;
 556     }
 557     /* define parameters */
 558     while ((sym = sym->next) != NULL) {
 559         type = &sym->type;
 560         size = type_size(type, &align);
 561         size = (size + 3) & ~3;
 562 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 563         /* structs are passed as pointer */
 564         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 565             size = 4;
 566         }
 567 #endif
 568         if (param_index < fastcall_nb_regs) {
 569             /* save FASTCALL register */
 570             loc -= 4;
 571             o(0x89);     /* movl */
 572             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 573             param_addr = loc;
 574         } else {
 575             param_addr = addr;
 576             addr += size;
 577         }
 578         sym_push(sym->v & ~SYM_FIELD, type,
 579                  VT_LOCAL | lvalue_type(type->t), param_addr);
 580         param_index++;
 581     }
 582     func_ret_sub = 0;
 583     /* pascal type call or fastcall ? */
 584     if (func_call == FUNC_STDCALL || func_call == FUNC_FASTCALLW)
 585         func_ret_sub = addr - 8;
 586 #ifndef TCC_TARGET_PE
 587     else if (func_vc)
 588         func_ret_sub = 4;
 589 #endif
 590
 591 #ifdef CONFIG_TCC_BCHECK
 592     /* leave some room for bound checking code */
 593     if (tcc_state->do_bounds_check) {
 594         func_bound_offset = lbounds_section->data_offset;
 595         func_bound_ind = ind;
 596         oad(0xb8, 0); /* lbound section pointer */
 597         oad(0xb8, 0); /* call to function */
 598     }
 599 #endif
 600 }
 601
 602 /* generate function epilog */
 603 ST_FUNC void gfunc_epilog(void)
 604 {
 605     addr_t v, saved_ind;
 606
 607 #ifdef CONFIG_TCC_BCHECK
 608     if (tcc_state->do_bounds_check
 609      && func_bound_offset != lbounds_section->data_offset) {
 610         addr_t saved_ind;
 611         addr_t *bounds_ptr;
 612         Sym *sym_data;
 613
 614         /* add end of table info */
 615         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 616         *bounds_ptr = 0;
 617
 618         /* generate bound local allocation */
 619         saved_ind = ind;
 620         ind = func_bound_ind;
 621         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 622                                func_bound_offset, lbounds_section->data_offset);
 623         greloc(cur_text_section, sym_data,
 624                ind + 1, R_386_32);
 625         oad(0xb8, 0); /* mov %eax, xxx */
 626         gen_static_call(TOK___bound_local_new);
 627         ind = saved_ind;
 628
 629         /* generate bound check local freeing */
 630         o(0x5250); /* save returned value, if any */
 631         greloc(cur_text_section, sym_data, ind + 1, R_386_32);
 632         oad(0xb8, 0); /* mov %eax, xxx */
 633         gen_static_call(TOK___bound_local_delete);
 634         o(0x585a); /* restore returned value, if any */
 635     }
 636 #endif
 637
 638     /* align local size to word & save local variables */
 639     v = (-loc + 3) & -4;
 640
 641 #if USE_EBX
 642     o(0x8b);
 643     gen_modrm(TREG_EBX, VT_LOCAL, NULL, -(v+4));
 644 #endif
 645
 646     o(0xc9); /* leave */
 647     if (func_ret_sub == 0) {
 648         o(0xc3); /* ret */
 649     } else {
 650         o(0xc2); /* ret n */
 651         g(func_ret_sub);
 652         g(func_ret_sub >> 8);
 653     }
 654     saved_ind = ind;
 655     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 656 #ifdef TCC_TARGET_PE
 657     if (v >= 4096) {
 658         oad(0xb8, v); /* mov stacksize, %eax */
 659         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 660     } else
 661 #endif
 662     {
 663         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 664         o(0xec81);  /* sub esp, stacksize */
 665         gen_le32(v);
 666 #ifdef TCC_TARGET_PE
 667         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 668 #endif
 669     }
 670     o(0x53 * USE_EBX); /* push ebx */
 671     ind = saved_ind;
 672 }
 673
 674 /* generate a jump to a label */
 675 ST_FUNC int gjmp(int t)
 676 {
 677     return gjmp2(0xe9, t);
 678 }
 679
 680 /* generate a jump to a fixed address */
 681 ST_FUNC void gjmp_addr(int a)
 682 {
 683     int r;
 684     r = a - ind - 2;
 685     if (r == (char)r) {
 686         g(0xeb);
 687         g(r);
 688     } else {
 689         oad(0xe9, a - ind - 5);
 690     }
 691 }
 692
 693 ST_FUNC void gtst_addr(int inv, int a)
 694 {
 695     int v = vtop->r & VT_VALMASK;
 696     if (v == VT_CMP) {
 697         inv ^= (vtop--)->c.i;
 698         a -= ind + 2;
 699         if (a == (char)a) {
 700             g(inv - 32);
 701             g(a);
 702         } else {
 703             g(0x0f);
 704             oad(inv - 16, a - 4);
 705         }
 706     } else if ((v & ~1) == VT_JMP) {
 707         if ((v & 1) != inv) {
 708             gjmp_addr(a);
 709             gsym(vtop->c.i);
 710         } else {
 711             gsym(vtop->c.i);
 712             o(0x05eb);
 713             gjmp_addr(a);
 714         }
 715         vtop--;
 716     }
 717 }
 718
 719 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 720 ST_FUNC int gtst(int inv, int t)
 721 {
 722     int v = vtop->r & VT_VALMASK;
 723     if (nocode_wanted) {
 724         ;
 725     } else if (v == VT_CMP) {
 726         /* fast case : can jump directly since flags are set */
 727         g(0x0f);
 728         t = gjmp2((vtop->c.i - 16) ^ inv, t);
 729     } else if (v == VT_JMP || v == VT_JMPI) {
 730         /* && or || optimization */
 731         if ((v & 1) == inv) {
 732             /* insert vtop->c jump list in t */
 733             uint32_t n1, n = vtop->c.i;
 734             if (n) {
 735                 while ((n1 = read32le(cur_text_section->data + n)))
 736                     n = n1;
 737                 write32le(cur_text_section->data + n, t);
 738                 t = vtop->c.i;
 739             }
 740         } else {
 741             t = gjmp(t);
 742             gsym(vtop->c.i);
 743         }
 744     }
 745     vtop--;
 746     return t;
 747 }
 748
 749 /* generate an integer binary operation */
 750 ST_FUNC void gen_opi(int op)
 751 {
 752     int r, fr, opc, c;
 753
 754     switch(op) {
 755     case '+':
 756     case TOK_ADDC1: /* add with carry generation */
 757         opc = 0;
 758     gen_op8:
 759         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 760             /* constant case */
 761             vswap();
 762             r = gv(RC_INT);
 763             vswap();
 764             c = vtop->c.i;
 765             if (c == (char)c) {
 766                 /* generate inc and dec for smaller code */
 767                 if (c==1 && opc==0 && op != TOK_ADDC1) {
 768                     o (0x40 | r); // inc
 769                 } else if (c==1 && opc==5 && op != TOK_SUBC1) {
 770                     o (0x48 | r); // dec
 771                 } else {
 772                     o(0x83);
 773                     o(0xc0 | (opc << 3) | r);
 774                     g(c);
 775                 }
 776             } else {
 777                 o(0x81);
 778                 oad(0xc0 | (opc << 3) | r, c);
 779             }
 780         } else {
 781             gv2(RC_INT, RC_INT);
 782             r = vtop[-1].r;
 783             fr = vtop[0].r;
 784             o((opc << 3) | 0x01);
 785             o(0xc0 + r + fr * 8);
 786         }
 787         vtop--;
 788         if (op >= TOK_ULT && op <= TOK_GT) {
 789             vtop->r = VT_CMP;
 790             vtop->c.i = op;
 791         }
 792         break;
 793     case '-':
 794     case TOK_SUBC1: /* sub with carry generation */
 795         opc = 5;
 796         goto gen_op8;
 797     case TOK_ADDC2: /* add with carry use */
 798         opc = 2;
 799         goto gen_op8;
 800     case TOK_SUBC2: /* sub with carry use */
 801         opc = 3;
 802         goto gen_op8;
 803     case '&':
 804         opc = 4;
 805         goto gen_op8;
 806     case '^':
 807         opc = 6;
 808         goto gen_op8;
 809     case '|':
 810         opc = 1;
 811         goto gen_op8;
 812     case '*':
 813         gv2(RC_INT, RC_INT);
 814         r = vtop[-1].r;
 815         fr = vtop[0].r;
 816         vtop--;
 817         o(0xaf0f); /* imul fr, r */
 818         o(0xc0 + fr + r * 8);
 819         break;
 820     case TOK_SHL:
 821         opc = 4;
 822         goto gen_shift;
 823     case TOK_SHR:
 824         opc = 5;
 825         goto gen_shift;
 826     case TOK_SAR:
 827         opc = 7;
 828     gen_shift:
 829         opc = 0xc0 | (opc << 3);
 830         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 831             /* constant case */
 832             vswap();
 833             r = gv(RC_INT);
 834             vswap();
 835             c = vtop->c.i & 0x1f;
 836             o(0xc1); /* shl/shr/sar $xxx, r */
 837             o(opc | r);
 838             g(c);
 839         } else {
 840             /* we generate the shift in ecx */
 841             gv2(RC_INT, RC_ECX);
 842             r = vtop[-1].r;
 843             o(0xd3); /* shl/shr/sar %cl, r */
 844             o(opc | r);
 845         }
 846         vtop--;
 847         break;
 848     case '/':
 849     case TOK_UDIV:
 850     case TOK_PDIV:
 851     case '%':
 852     case TOK_UMOD:
 853     case TOK_UMULL:
 854         /* first operand must be in eax */
 855         /* XXX: need better constraint for second operand */
 856         gv2(RC_EAX, RC_ECX);
 857         r = vtop[-1].r;
 858         fr = vtop[0].r;
 859         vtop--;
 860         save_reg(TREG_EDX);
 861         /* save EAX too if used otherwise */
 862         save_reg_upstack(TREG_EAX, 1);
 863         if (op == TOK_UMULL) {
 864             o(0xf7); /* mul fr */
 865             o(0xe0 + fr);
 866             vtop->r2 = TREG_EDX;
 867             r = TREG_EAX;
 868         } else {
 869             if (op == TOK_UDIV || op == TOK_UMOD) {
 870                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 871                 o(0xf0 + fr);
 872             } else {
 873                 o(0xf799); /* cltd, idiv fr, %eax */
 874                 o(0xf8 + fr);
 875             }
 876             if (op == '%' || op == TOK_UMOD)
 877                 r = TREG_EDX;
 878             else
 879                 r = TREG_EAX;
 880         }
 881         vtop->r = r;
 882         break;
 883     default:
 884         opc = 7;
 885         goto gen_op8;
 886     }
 887 }
 888
 889 /* generate a floating point operation 'v = t1 op t2' instruction. The
 890    two operands are guaranteed to have the same floating point type */
 891 /* XXX: need to use ST1 too */
 892 ST_FUNC void gen_opf(int op)
 893 {
 894     int a, ft, fc, swapped, r;
 895
 896     /* convert constants to memory references */
 897     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 898         vswap();
 899         gv(RC_FLOAT);
 900         vswap();
 901     }
 902     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 903         gv(RC_FLOAT);
 904
 905     /* must put at least one value in the floating point register */
 906     if ((vtop[-1].r & VT_LVAL) &&
 907         (vtop[0].r & VT_LVAL)) {
 908         vswap();
 909         gv(RC_FLOAT);
 910         vswap();
 911     }
 912     swapped = 0;
 913     /* swap the stack if needed so that t1 is the register and t2 is
 914        the memory reference */
 915     if (vtop[-1].r & VT_LVAL) {
 916         vswap();
 917         swapped = 1;
 918     }
 919     if (op >= TOK_ULT && op <= TOK_GT) {
 920         /* load on stack second operand */
 921         load(TREG_ST0, vtop);
 922         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 923         if (op == TOK_GE || op == TOK_GT)
 924             swapped = !swapped;
 925         else if (op == TOK_EQ || op == TOK_NE)
 926             swapped = 0;
 927         if (swapped)
 928             o(0xc9d9); /* fxch %st(1) */
 929         if (op == TOK_EQ || op == TOK_NE)
 930             o(0xe9da); /* fucompp */
 931         else
 932             o(0xd9de); /* fcompp */
 933         o(0xe0df); /* fnstsw %ax */
 934         if (op == TOK_EQ) {
 935             o(0x45e480); /* and $0x45, %ah */
 936             o(0x40fC80); /* cmp $0x40, %ah */
 937         } else if (op == TOK_NE) {
 938             o(0x45e480); /* and $0x45, %ah */
 939             o(0x40f480); /* xor $0x40, %ah */
 940             op = TOK_NE;
 941         } else if (op == TOK_GE || op == TOK_LE) {
 942             o(0x05c4f6); /* test $0x05, %ah */
 943             op = TOK_EQ;
 944         } else {
 945             o(0x45c4f6); /* test $0x45, %ah */
 946             op = TOK_EQ;
 947         }
 948         vtop--;
 949         vtop->r = VT_CMP;
 950         vtop->c.i = op;
 951     } else {
 952         /* no memory reference possible for long double operations */
 953         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 954             load(TREG_ST0, vtop);
 955             swapped = !swapped;
 956         }
 957
 958         switch(op) {
 959         default:
 960         case '+':
 961             a = 0;
 962             break;
 963         case '-':
 964             a = 4;
 965             if (swapped)
 966                 a++;
 967             break;
 968         case '*':
 969             a = 1;
 970             break;
 971         case '/':
 972             a = 6;
 973             if (swapped)
 974                 a++;
 975             break;
 976         }
 977         ft = vtop->type.t;
 978         fc = vtop->c.i;
 979         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 980             o(0xde); /* fxxxp %st, %st(1) */
 981             o(0xc1 + (a << 3));
 982         } else {
 983             /* if saved lvalue, then we must reload it */
 984             r = vtop->r;
 985             if ((r & VT_VALMASK) == VT_LLOCAL) {
 986                 SValue v1;
 987                 r = get_reg(RC_INT);
 988                 v1.type.t = VT_INT;
 989                 v1.r = VT_LOCAL | VT_LVAL;
 990                 v1.c.i = fc;
 991                 load(r, &v1);
 992                 fc = 0;
 993             }
 994
 995             if ((ft & VT_BTYPE) == VT_DOUBLE)
 996                 o(0xdc);
 997             else
 998                 o(0xd8);
 999             gen_modrm(a, r, vtop->sym, fc);
1000         }
1001         vtop--;
1002     }
1003 }
1004
1005 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1006    and 'long long' cases. */
1007 ST_FUNC void gen_cvt_itof(int t)
1008 {
1009     save_reg(TREG_ST0);
1010     gv(RC_INT);
1011     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
1012         /* signed long long to float/double/long double (unsigned case
1013            is handled generically) */
1014         o(0x50 + vtop->r2); /* push r2 */
1015         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1016         o(0x242cdf); /* fildll (%esp) */
1017         o(0x08c483); /* add $8, %esp */
1018     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
1019                (VT_INT | VT_UNSIGNED)) {
1020         /* unsigned int to float/double/long double */
1021         o(0x6a); /* push $0 */
1022         g(0x00);
1023         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1024         o(0x242cdf); /* fildll (%esp) */
1025         o(0x08c483); /* add $8, %esp */
1026     } else {
1027         /* int to float/double/long double */
1028         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1029         o(0x2404db); /* fildl (%esp) */
1030         o(0x04c483); /* add $4, %esp */
1031     }
1032     vtop->r = TREG_ST0;
1033 }
1034
1035 /* convert fp to int 't' type */
1036 ST_FUNC void gen_cvt_ftoi(int t)
1037 {
1038     int bt = vtop->type.t & VT_BTYPE;
1039     if (bt == VT_FLOAT)
1040         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1041     else if (bt == VT_LDOUBLE)
1042         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1043     else
1044         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1045     vswap();
1046     gfunc_call(1);
1047     vpushi(0);
1048     vtop->r = REG_IRET;
1049     vtop->r2 = REG_LRET;
1050 }
1051
1052 /* convert from one floating point type to another */
1053 ST_FUNC void gen_cvt_ftof(int t)
1054 {
1055     /* all we have to do on i386 is to put the float in a register */
1056     gv(RC_FLOAT);
1057 }
1058
1059 /* computed goto support */
1060 ST_FUNC void ggoto(void)
1061 {
1062     gcall_or_jmp(1);
1063     vtop--;
1064 }
1065
1066 /* bound check support functions */
1067 #ifdef CONFIG_TCC_BCHECK
1068
1069 /* generate a bounded pointer addition */
1070 ST_FUNC void gen_bounded_ptr_add(void)
1071 {
1072     /* prepare fast i386 function call (args in eax and edx) */
1073     gv2(RC_EAX, RC_EDX);
1074     /* save all temporary registers */
1075     vtop -= 2;
1076     save_regs(0);
1077     /* do a fast function call */
1078     gen_static_call(TOK___bound_ptr_add);
1079     /* returned pointer is in eax */
1080     vtop++;
1081     vtop->r = TREG_EAX | VT_BOUNDED;
1082     /* address of bounding function call point */
1083     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1084 }
1085
1086 /* patch pointer addition in vtop so that pointer dereferencing is
1087    also tested */
1088 ST_FUNC void gen_bounded_ptr_deref(void)
1089 {
1090     addr_t func;
1091     int  size, align;
1092     Elf32_Rel *rel;
1093     Sym *sym;
1094
1095     size = 0;
1096     /* XXX: put that code in generic part of tcc */
1097     if (!is_float(vtop->type.t)) {
1098         if (vtop->r & VT_LVAL_BYTE)
1099             size = 1;
1100         else if (vtop->r & VT_LVAL_SHORT)
1101             size = 2;
1102     }
1103     if (!size)
1104         size = type_size(&vtop->type, &align);
1105     switch(size) {
1106     case  1: func = TOK___bound_ptr_indir1; break;
1107     case  2: func = TOK___bound_ptr_indir2; break;
1108     case  4: func = TOK___bound_ptr_indir4; break;
1109     case  8: func = TOK___bound_ptr_indir8; break;
1110     case 12: func = TOK___bound_ptr_indir12; break;
1111     case 16: func = TOK___bound_ptr_indir16; break;
1112     default:
1113         tcc_error("unhandled size when dereferencing bounded pointer");
1114         func = 0;
1115         break;
1116     }
1117
1118     /* patch relocation */
1119     /* XXX: find a better solution ? */
1120     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1121     sym = external_global_sym(func, &func_old_type, 0);
1122     if (!sym->c)
1123         put_extern_sym(sym, NULL, 0, 0);
1124     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1125 }
1126 #endif
1127
1128 /* Save the stack pointer onto the stack */
1129 ST_FUNC void gen_vla_sp_save(int addr) {
1130     /* mov %esp,addr(%ebp)*/
1131     o(0x89);
1132     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1133 }
1134
1135 /* Restore the SP from a location on the stack */
1136 ST_FUNC void gen_vla_sp_restore(int addr) {
1137     o(0x8b);
1138     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1139 }
1140
1141 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1142 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1143 #ifdef TCC_TARGET_PE
1144     /* alloca does more than just adjust %rsp on Windows */
1145     vpush_global_sym(&func_old_type, TOK_alloca);
1146     vswap(); /* Move alloca ref past allocation size */
1147     gfunc_call(1);
1148 #else
1149     int r;
1150     r = gv(RC_INT); /* allocation size */
1151     /* sub r,%rsp */
1152     o(0x2b);
1153     o(0xe0 | r);
1154     /* We align to 16 bytes rather than align */
1155     /* and ~15, %esp */
1156     o(0xf0e483);
1157     vpop();
1158 #endif
1159 }
1160
1161 /* end of X86 code generator */
1162 /*************************************************************/
1163 #endif
1164 /*************************************************************/