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
  27 /* a register can belong to several classes. The classes must be
  28    sorted from more general to more precise (see gv2() code which does
  29    assumptions on it). */
  30 #define RC_INT     0x0001 /* generic integer register */
  31 #define RC_FLOAT   0x0002 /* generic float register */
  32 #define RC_EAX     0x0004
  33 #define RC_ST0     0x0008
  34 #define RC_ECX     0x0010
  35 #define RC_EDX     0x0020
  36 #define RC_EBX     0x0040
  37
  38 #define RC_IRET    RC_EAX /* function return: integer register */
  39 #define RC_LRET    RC_EDX /* function return: second integer register */
  40 #define RC_FRET    RC_ST0 /* function return: float register */
  41
  42 /* pretty names for the registers */
  43 enum {
  44     TREG_EAX = 0,
  45     TREG_ECX,
  46     TREG_EDX,
  47     TREG_EBX,
  48     TREG_ST0,
  49     TREG_ESP = 4
  50 };
  51
  52 /* return registers for function */
  53 #define REG_IRET TREG_EAX /* single word int return register */
  54 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  55 #define REG_FRET TREG_ST0 /* float return register */
  56
  57 /* defined if function parameters must be evaluated in reverse order */
  58 #define INVERT_FUNC_PARAMS
  59
  60 /* defined if structures are passed as pointers. Otherwise structures
  61    are directly pushed on stack. */
  62 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  63
  64 /* pointer size, in bytes */
  65 #define PTR_SIZE 4
  66
  67 /* long double size and alignment, in bytes */
  68 #define LDOUBLE_SIZE  12
  69 #define LDOUBLE_ALIGN 4
  70 /* maximum alignment (for aligned attribute support) */
  71 #define MAX_ALIGN     8
  72
  73
  74 #define psym oad
  75
  76 /******************************************************/
  77 #else /* ! TARGET_DEFS_ONLY */
  78 /******************************************************/
  79 #include "tcc.h"
  80
  81 /* define to 1/0 to [not] have EBX as 4th register */
  82 #define USE_EBX 1
  83
  84 ST_DATA const int reg_classes[NB_REGS] = {
  85     /* eax */ RC_INT | RC_EAX,
  86     /* ecx */ RC_INT | RC_ECX,
  87     /* edx */ RC_INT | RC_EDX,
  88     /* ebx */ (RC_INT | RC_EBX) * USE_EBX,
  89     /* st0 */ RC_FLOAT | RC_ST0,
  90 };
  91
  92 static unsigned long func_sub_sp_offset;
  93 static int func_ret_sub;
  94 #ifdef CONFIG_TCC_BCHECK
  95 static addr_t func_bound_offset;
  96 static unsigned long func_bound_ind;
  97 #endif
  98
  99 /* XXX: make it faster ? */
 100 ST_FUNC void g(int c)
 101 {
 102     int ind1;
 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 /* psym is used to put an instruction with a data field which is a
 149    reference to a symbol. It is in fact the same as oad ! */
 150 #define psym oad
 151
 152 /* instruction + 4 bytes data. Return the address of the data */
 153 ST_FUNC int oad(int c, int s)
 154 {
 155     int ind1;
 156
 157     o(c);
 158     ind1 = ind + 4;
 159     if (ind1 > cur_text_section->data_allocated)
 160         section_realloc(cur_text_section, ind1);
 161     write32le(cur_text_section->data + ind, s);
 162     s = ind;
 163     ind = ind1;
 164     return s;
 165 }
 166
 167 /* output constant with relocation if 'r & VT_SYM' is true */
 168 ST_FUNC void gen_addr32(int r, Sym *sym, long c)
 169 {
 170     if (r & VT_SYM)
 171         greloc(cur_text_section, sym, ind, R_386_32);
 172     gen_le32(c);
 173 }
 174
 175 ST_FUNC void gen_addrpc32(int r, Sym *sym, long c)
 176 {
 177     if (r & VT_SYM)
 178         greloc(cur_text_section, sym, ind, R_386_PC32);
 179     gen_le32(c - 4);
 180 }
 181
 182 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 183    opcode bits */
 184 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 185 {
 186     op_reg = op_reg << 3;
 187     if ((r & VT_VALMASK) == VT_CONST) {
 188         /* constant memory reference */
 189         o(0x05 | op_reg);
 190         gen_addr32(r, sym, c);
 191     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 192         /* currently, we use only ebp as base */
 193         if (c == (char)c) {
 194             /* short reference */
 195             o(0x45 | op_reg);
 196             g(c);
 197         } else {
 198             oad(0x85 | op_reg, c);
 199         }
 200     } else {
 201         g(0x00 | op_reg | (r & VT_VALMASK));
 202     }
 203 }
 204
 205 /* load 'r' from value 'sv' */
 206 ST_FUNC void load(int r, SValue *sv)
 207 {
 208     int v, t, ft, fc, fr;
 209     SValue v1;
 210
 211 #ifdef TCC_TARGET_PE
 212     SValue v2;
 213     sv = pe_getimport(sv, &v2);
 214 #endif
 215
 216     fr = sv->r;
 217     ft = sv->type.t;
 218     fc = sv->c.i;
 219
 220     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 221
 222     v = fr & VT_VALMASK;
 223     if (fr & VT_LVAL) {
 224         if (v == VT_LLOCAL) {
 225             v1.type.t = VT_INT;
 226             v1.r = VT_LOCAL | VT_LVAL;
 227             v1.c.i = fc;
 228             fr = r;
 229             if (!(reg_classes[fr] & RC_INT))
 230                 fr = get_reg(RC_INT);
 231             load(fr, &v1);
 232         }
 233         if ((ft & VT_BTYPE) == VT_FLOAT) {
 234             o(0xd9); /* flds */
 235             r = 0;
 236         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 237             o(0xdd); /* fldl */
 238             r = 0;
 239         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 240             o(0xdb); /* fldt */
 241             r = 5;
 242         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 243             o(0xbe0f);   /* movsbl */
 244         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 245             o(0xb60f);   /* movzbl */
 246         } else if ((ft & VT_TYPE) == VT_SHORT) {
 247             o(0xbf0f);   /* movswl */
 248         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 249             o(0xb70f);   /* movzwl */
 250         } else {
 251             o(0x8b);     /* movl */
 252         }
 253         gen_modrm(r, fr, sv->sym, fc);
 254     } else {
 255         if (v == VT_CONST) {
 256             o(0xb8 + r); /* mov $xx, r */
 257             gen_addr32(fr, sv->sym, fc);
 258         } else if (v == VT_LOCAL) {
 259             if (fc) {
 260                 o(0x8d); /* lea xxx(%ebp), r */
 261                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 262             } else {
 263                 o(0x89);
 264                 o(0xe8 + r); /* mov %ebp, r */
 265             }
 266         } else if (v == VT_CMP) {
 267             oad(0xb8 + r, 0); /* mov $0, r */
 268             o(0x0f); /* setxx %br */
 269             o(fc);
 270             o(0xc0 + r);
 271         } else if (v == VT_JMP || v == VT_JMPI) {
 272             t = v & 1;
 273             oad(0xb8 + r, t); /* mov $1, r */
 274             o(0x05eb); /* jmp after */
 275             gsym(fc);
 276             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 277         } else if (v != r) {
 278             o(0x89);
 279             o(0xc0 + r + v * 8); /* mov v, r */
 280         }
 281     }
 282 }
 283
 284 /* store register 'r' in lvalue 'v' */
 285 ST_FUNC void store(int r, SValue *v)
 286 {
 287     int fr, bt, ft, fc;
 288
 289 #ifdef TCC_TARGET_PE
 290     SValue v2;
 291     v = pe_getimport(v, &v2);
 292 #endif
 293
 294     ft = v->type.t;
 295     fc = v->c.i;
 296     fr = v->r & VT_VALMASK;
 297     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 298     bt = ft & VT_BTYPE;
 299     /* XXX: incorrect if float reg to reg */
 300     if (bt == VT_FLOAT) {
 301         o(0xd9); /* fsts */
 302         r = 2;
 303     } else if (bt == VT_DOUBLE) {
 304         o(0xdd); /* fstpl */
 305         r = 2;
 306     } else if (bt == VT_LDOUBLE) {
 307         o(0xc0d9); /* fld %st(0) */
 308         o(0xdb); /* fstpt */
 309         r = 7;
 310     } else {
 311         if (bt == VT_SHORT)
 312             o(0x66);
 313         if (bt == VT_BYTE || bt == VT_BOOL)
 314             o(0x88);
 315         else
 316             o(0x89);
 317     }
 318     if (fr == VT_CONST ||
 319         fr == VT_LOCAL ||
 320         (v->r & VT_LVAL)) {
 321         gen_modrm(r, v->r, v->sym, fc);
 322     } else if (fr != r) {
 323         o(0xc0 + fr + r * 8); /* mov r, fr */
 324     }
 325 }
 326
 327 static void gadd_sp(int val)
 328 {
 329     if (val == (char)val) {
 330         o(0xc483);
 331         g(val);
 332     } else {
 333         oad(0xc481, val); /* add $xxx, %esp */
 334     }
 335 }
 336
 337 static void gen_static_call(int v)
 338 {
 339     Sym *sym;
 340
 341     sym = external_global_sym(v, &func_old_type, 0);
 342     oad(0xe8, -4);
 343     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 344 }
 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) {
 351         int rt;
 352         /* constant case */
 353         if (vtop->r & VT_SYM) {
 354             /* relocation case */
 355             greloc(cur_text_section, vtop->sym,
 356                    ind + 1, R_386_PC32);
 357         } else {
 358             /* put an empty PC32 relocation */
 359             put_elf_reloc(symtab_section, cur_text_section,
 360                           ind + 1, R_386_PC32, 0);
 361         }
 362         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 363         /* extend the return value to the whole register if necessary
 364            visual studio and gcc do not always set the whole eax register
 365            when assigning the return value of a function  */
 366         rt = vtop->type.ref->type.t;
 367         switch (rt & VT_BTYPE) {
 368             case VT_BYTE:
 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     } else {
 388         /* otherwise, indirect call */
 389         r = gv(RC_INT);
 390         o(0xff); /* call/jmp *r */
 391         o(0xd0 + r + (is_jmp << 4));
 392     }
 393 }
 394
 395 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 396 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 397
 398 /* Return the number of registers needed to return the struct, or 0 if
 399    returning via struct pointer. */
 400 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 401 {
 402 #ifdef TCC_TARGET_PE
 403     int size, align;
 404
 405     *ret_align = 1; // Never have to re-align return values for x86
 406     *regsize = 4;
 407     size = type_size(vt, &align);
 408     if (size > 8) {
 409         return 0;
 410     } else if (size > 4) {
 411         ret->ref = NULL;
 412         ret->t = VT_LLONG;
 413         return 1;
 414     } else {
 415         ret->ref = NULL;
 416         ret->t = VT_INT;
 417         return 1;
 418     }
 419 #else
 420     *ret_align = 1; // Never have to re-align return values for x86
 421     return 0;
 422 #endif
 423 }
 424
 425 /* Generate function call. The function address is pushed first, then
 426    all the parameters in call order. This functions pops all the
 427    parameters and the function address. */
 428 ST_FUNC void gfunc_call(int nb_args)
 429 {
 430     int size, align, r, args_size, i, func_call;
 431     Sym *func_sym;
 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->a.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)
 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(CType *func_type)
 522 {
 523     int addr, align, size, func_call, fastcall_nb_regs;
 524     int param_index, param_addr;
 525     uint8_t *fastcall_regs_ptr;
 526     Sym *sym;
 527     CType *type;
 528
 529     sym = func_type->ref;
 530     func_call = sym->a.func_call;
 531     addr = 8;
 532     loc = 0;
 533     func_vc = 0;
 534
 535     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 536         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 537         fastcall_regs_ptr = fastcall_regs;
 538     } else if (func_call == FUNC_FASTCALLW) {
 539         fastcall_nb_regs = 2;
 540         fastcall_regs_ptr = fastcallw_regs;
 541     } else {
 542         fastcall_nb_regs = 0;
 543         fastcall_regs_ptr = NULL;
 544     }
 545     param_index = 0;
 546
 547     ind += FUNC_PROLOG_SIZE;
 548     func_sub_sp_offset = ind;
 549     /* if the function returns a structure, then add an
 550        implicit pointer parameter */
 551     func_vt = sym->type;
 552     func_var = (sym->c == FUNC_ELLIPSIS);
 553 #ifdef TCC_TARGET_PE
 554     size = type_size(&func_vt,&align);
 555     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 556 #else
 557     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 558 #endif
 559         /* XXX: fastcall case ? */
 560         func_vc = addr;
 561         addr += 4;
 562         param_index++;
 563     }
 564     /* define parameters */
 565     while ((sym = sym->next) != NULL) {
 566         type = &sym->type;
 567         size = type_size(type, &align);
 568         size = (size + 3) & ~3;
 569 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 570         /* structs are passed as pointer */
 571         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 572             size = 4;
 573         }
 574 #endif
 575         if (param_index < fastcall_nb_regs) {
 576             /* save FASTCALL register */
 577             loc -= 4;
 578             o(0x89);     /* movl */
 579             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 580             param_addr = loc;
 581         } else {
 582             param_addr = addr;
 583             addr += size;
 584         }
 585         sym_push(sym->v & ~SYM_FIELD, type,
 586                  VT_LOCAL | lvalue_type(type->t), param_addr);
 587         param_index++;
 588     }
 589     func_ret_sub = 0;
 590     /* pascal type call ? */
 591     if (func_call == FUNC_STDCALL)
 592         func_ret_sub = addr - 8;
 593 #ifndef TCC_TARGET_PE
 594     else if (func_vc)
 595         func_ret_sub = 4;
 596 #endif
 597
 598 #ifdef CONFIG_TCC_BCHECK
 599     /* leave some room for bound checking code */
 600     if (tcc_state->do_bounds_check) {
 601         func_bound_offset = lbounds_section->data_offset;
 602         func_bound_ind = ind;
 603         oad(0xb8, 0); /* lbound section pointer */
 604         oad(0xb8, 0); /* call to function */
 605     }
 606 #endif
 607 }
 608
 609 /* generate function epilog */
 610 ST_FUNC void gfunc_epilog(void)
 611 {
 612     addr_t v, saved_ind;
 613
 614 #ifdef CONFIG_TCC_BCHECK
 615     if (tcc_state->do_bounds_check
 616      && func_bound_offset != lbounds_section->data_offset) {
 617         addr_t saved_ind;
 618         addr_t *bounds_ptr;
 619         Sym *sym_data;
 620
 621         /* add end of table info */
 622         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 623         *bounds_ptr = 0;
 624
 625         /* generate bound local allocation */
 626         saved_ind = ind;
 627         ind = func_bound_ind;
 628         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 629                                func_bound_offset, lbounds_section->data_offset);
 630         greloc(cur_text_section, sym_data,
 631                ind + 1, R_386_32);
 632         oad(0xb8, 0); /* mov %eax, xxx */
 633         gen_static_call(TOK___bound_local_new);
 634         ind = saved_ind;
 635
 636         /* generate bound check local freeing */
 637         o(0x5250); /* save returned value, if any */
 638         greloc(cur_text_section, sym_data, ind + 1, R_386_32);
 639         oad(0xb8, 0); /* mov %eax, xxx */
 640         gen_static_call(TOK___bound_local_delete);
 641         o(0x585a); /* restore returned value, if any */
 642     }
 643 #endif
 644     o(0x5b * USE_EBX); /* pop ebx */
 645     o(0xc9); /* leave */
 646     if (func_ret_sub == 0) {
 647         o(0xc3); /* ret */
 648     } else {
 649         o(0xc2); /* ret n */
 650         g(func_ret_sub);
 651         g(func_ret_sub >> 8);
 652     }
 653     /* align local size to word & save local variables */
 654
 655     v = (-loc + 3) & -4;
 656     saved_ind = ind;
 657     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 658 #ifdef TCC_TARGET_PE
 659     if (v >= 4096) {
 660         oad(0xb8, v); /* mov stacksize, %eax */
 661         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 662     } else
 663 #endif
 664     {
 665         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 666         o(0xec81);  /* sub esp, stacksize */
 667         gen_le32(v);
 668 #ifdef TCC_TARGET_PE
 669         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 670 #endif
 671     }
 672     o(0x53 * USE_EBX); /* push ebx */
 673     ind = saved_ind;
 674 }
 675
 676 /* generate a jump to a label */
 677 ST_FUNC int gjmp(int t)
 678 {
 679     return psym(0xe9, t);
 680 }
 681
 682 /* generate a jump to a fixed address */
 683 ST_FUNC void gjmp_addr(int a)
 684 {
 685     int r;
 686     r = a - ind - 2;
 687     if (r == (char)r) {
 688         g(0xeb);
 689         g(r);
 690     } else {
 691         oad(0xe9, a - ind - 5);
 692     }
 693 }
 694
 695 ST_FUNC void gtst_addr(int inv, int a)
 696 {
 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 }
 707
 708 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 709 ST_FUNC int gtst(int inv, int t)
 710 {
 711     int v = vtop->r & VT_VALMASK;
 712     if (v == VT_CMP) {
 713         /* fast case : can jump directly since flags are set */
 714         g(0x0f);
 715         t = psym((vtop->c.i - 16) ^ inv, t);
 716     } else if (v == VT_JMP || v == VT_JMPI) {
 717         /* && or || optimization */
 718         if ((v & 1) == inv) {
 719             /* insert vtop->c jump list in t */
 720             uint32_t n1, n = vtop->c.i;
 721             if (n) {
 722                 while ((n1 = read32le(cur_text_section->data + n)))
 723                     n = n1;
 724                 write32le(cur_text_section->data + n, t);
 725                 t = vtop->c.i;
 726             }
 727         } else {
 728             t = gjmp(t);
 729             gsym(vtop->c.i);
 730         }
 731     }
 732     vtop--;
 733     return t;
 734 }
 735
 736 /* generate an integer binary operation */
 737 ST_FUNC void gen_opi(int op)
 738 {
 739     int r, fr, opc, c;
 740
 741     switch(op) {
 742     case '+':
 743     case TOK_ADDC1: /* add with carry generation */
 744         opc = 0;
 745     gen_op8:
 746         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 747             /* constant case */
 748             vswap();
 749             r = gv(RC_INT);
 750             vswap();
 751             c = vtop->c.i;
 752             if (c == (char)c) {
 753                 /* generate inc and dec for smaller code */
 754                 if (c==1 && opc==0 && op != TOK_ADDC1) {
 755                     o (0x40 | r); // inc
 756                 } else if (c==1 && opc==5 && op != TOK_SUBC1) {
 757                     o (0x48 | r); // dec
 758                 } else {
 759                     o(0x83);
 760                     o(0xc0 | (opc << 3) | r);
 761                     g(c);
 762                 }
 763             } else {
 764                 o(0x81);
 765                 oad(0xc0 | (opc << 3) | r, c);
 766             }
 767         } else {
 768             gv2(RC_INT, RC_INT);
 769             r = vtop[-1].r;
 770             fr = vtop[0].r;
 771             o((opc << 3) | 0x01);
 772             o(0xc0 + r + fr * 8);
 773         }
 774         vtop--;
 775         if (op >= TOK_ULT && op <= TOK_GT) {
 776             vtop->r = VT_CMP;
 777             vtop->c.i = op;
 778         }
 779         break;
 780     case '-':
 781     case TOK_SUBC1: /* sub with carry generation */
 782         opc = 5;
 783         goto gen_op8;
 784     case TOK_ADDC2: /* add with carry use */
 785         opc = 2;
 786         goto gen_op8;
 787     case TOK_SUBC2: /* sub with carry use */
 788         opc = 3;
 789         goto gen_op8;
 790     case '&':
 791         opc = 4;
 792         goto gen_op8;
 793     case '^':
 794         opc = 6;
 795         goto gen_op8;
 796     case '|':
 797         opc = 1;
 798         goto gen_op8;
 799     case '*':
 800         gv2(RC_INT, RC_INT);
 801         r = vtop[-1].r;
 802         fr = vtop[0].r;
 803         vtop--;
 804         o(0xaf0f); /* imul fr, r */
 805         o(0xc0 + fr + r * 8);
 806         break;
 807     case TOK_SHL:
 808         opc = 4;
 809         goto gen_shift;
 810     case TOK_SHR:
 811         opc = 5;
 812         goto gen_shift;
 813     case TOK_SAR:
 814         opc = 7;
 815     gen_shift:
 816         opc = 0xc0 | (opc << 3);
 817         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 818             /* constant case */
 819             vswap();
 820             r = gv(RC_INT);
 821             vswap();
 822             c = vtop->c.i & 0x1f;
 823             o(0xc1); /* shl/shr/sar $xxx, r */
 824             o(opc | r);
 825             g(c);
 826         } else {
 827             /* we generate the shift in ecx */
 828             gv2(RC_INT, RC_ECX);
 829             r = vtop[-1].r;
 830             o(0xd3); /* shl/shr/sar %cl, r */
 831             o(opc | r);
 832         }
 833         vtop--;
 834         break;
 835     case '/':
 836     case TOK_UDIV:
 837     case TOK_PDIV:
 838     case '%':
 839     case TOK_UMOD:
 840     case TOK_UMULL:
 841         /* first operand must be in eax */
 842         /* XXX: need better constraint for second operand */
 843         gv2(RC_EAX, RC_ECX);
 844         r = vtop[-1].r;
 845         fr = vtop[0].r;
 846         vtop--;
 847         save_reg(TREG_EDX);
 848         /* save EAX too if used otherwise */
 849         save_reg_upstack(TREG_EAX, 1);
 850         if (op == TOK_UMULL) {
 851             o(0xf7); /* mul fr */
 852             o(0xe0 + fr);
 853             vtop->r2 = TREG_EDX;
 854             r = TREG_EAX;
 855         } else {
 856             if (op == TOK_UDIV || op == TOK_UMOD) {
 857                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 858                 o(0xf0 + fr);
 859             } else {
 860                 o(0xf799); /* cltd, idiv fr, %eax */
 861                 o(0xf8 + fr);
 862             }
 863             if (op == '%' || op == TOK_UMOD)
 864                 r = TREG_EDX;
 865             else
 866                 r = TREG_EAX;
 867         }
 868         vtop->r = r;
 869         break;
 870     default:
 871         opc = 7;
 872         goto gen_op8;
 873     }
 874 }
 875
 876 /* generate a floating point operation 'v = t1 op t2' instruction. The
 877    two operands are guaranted to have the same floating point type */
 878 /* XXX: need to use ST1 too */
 879 ST_FUNC void gen_opf(int op)
 880 {
 881     int a, ft, fc, swapped, r;
 882
 883     /* convert constants to memory references */
 884     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 885         vswap();
 886         gv(RC_FLOAT);
 887         vswap();
 888     }
 889     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 890         gv(RC_FLOAT);
 891
 892     /* must put at least one value in the floating point register */
 893     if ((vtop[-1].r & VT_LVAL) &&
 894         (vtop[0].r & VT_LVAL)) {
 895         vswap();
 896         gv(RC_FLOAT);
 897         vswap();
 898     }
 899     swapped = 0;
 900     /* swap the stack if needed so that t1 is the register and t2 is
 901        the memory reference */
 902     if (vtop[-1].r & VT_LVAL) {
 903         vswap();
 904         swapped = 1;
 905     }
 906     if (op >= TOK_ULT && op <= TOK_GT) {
 907         /* load on stack second operand */
 908         load(TREG_ST0, vtop);
 909         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 910         if (op == TOK_GE || op == TOK_GT)
 911             swapped = !swapped;
 912         else if (op == TOK_EQ || op == TOK_NE)
 913             swapped = 0;
 914         if (swapped)
 915             o(0xc9d9); /* fxch %st(1) */
 916         if (op == TOK_EQ || op == TOK_NE)
 917             o(0xe9da); /* fucompp */
 918         else
 919             o(0xd9de); /* fcompp */
 920         o(0xe0df); /* fnstsw %ax */
 921         if (op == TOK_EQ) {
 922             o(0x45e480); /* and $0x45, %ah */
 923             o(0x40fC80); /* cmp $0x40, %ah */
 924         } else if (op == TOK_NE) {
 925             o(0x45e480); /* and $0x45, %ah */
 926             o(0x40f480); /* xor $0x40, %ah */
 927             op = TOK_NE;
 928         } else if (op == TOK_GE || op == TOK_LE) {
 929             o(0x05c4f6); /* test $0x05, %ah */
 930             op = TOK_EQ;
 931         } else {
 932             o(0x45c4f6); /* test $0x45, %ah */
 933             op = TOK_EQ;
 934         }
 935         vtop--;
 936         vtop->r = VT_CMP;
 937         vtop->c.i = op;
 938     } else {
 939         /* no memory reference possible for long double operations */
 940         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 941             load(TREG_ST0, vtop);
 942             swapped = !swapped;
 943         }
 944
 945         switch(op) {
 946         default:
 947         case '+':
 948             a = 0;
 949             break;
 950         case '-':
 951             a = 4;
 952             if (swapped)
 953                 a++;
 954             break;
 955         case '*':
 956             a = 1;
 957             break;
 958         case '/':
 959             a = 6;
 960             if (swapped)
 961                 a++;
 962             break;
 963         }
 964         ft = vtop->type.t;
 965         fc = vtop->c.i;
 966         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 967             o(0xde); /* fxxxp %st, %st(1) */
 968             o(0xc1 + (a << 3));
 969         } else {
 970             /* if saved lvalue, then we must reload it */
 971             r = vtop->r;
 972             if ((r & VT_VALMASK) == VT_LLOCAL) {
 973                 SValue v1;
 974                 r = get_reg(RC_INT);
 975                 v1.type.t = VT_INT;
 976                 v1.r = VT_LOCAL | VT_LVAL;
 977                 v1.c.i = fc;
 978                 load(r, &v1);
 979                 fc = 0;
 980             }
 981
 982             if ((ft & VT_BTYPE) == VT_DOUBLE)
 983                 o(0xdc);
 984             else
 985                 o(0xd8);
 986             gen_modrm(a, r, vtop->sym, fc);
 987         }
 988         vtop--;
 989     }
 990 }
 991
 992 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 993    and 'long long' cases. */
 994 ST_FUNC void gen_cvt_itof(int t)
 995 {
 996     save_reg(TREG_ST0);
 997     gv(RC_INT);
 998     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 999         /* signed long long to float/double/long double (unsigned case
1000            is handled generically) */
1001         o(0x50 + vtop->r2); /* push r2 */
1002         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1003         o(0x242cdf); /* fildll (%esp) */
1004         o(0x08c483); /* add $8, %esp */
1005     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
1006                (VT_INT | VT_UNSIGNED)) {
1007         /* unsigned int to float/double/long double */
1008         o(0x6a); /* push $0 */
1009         g(0x00);
1010         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1011         o(0x242cdf); /* fildll (%esp) */
1012         o(0x08c483); /* add $8, %esp */
1013     } else {
1014         /* int to float/double/long double */
1015         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1016         o(0x2404db); /* fildl (%esp) */
1017         o(0x04c483); /* add $4, %esp */
1018     }
1019     vtop->r = TREG_ST0;
1020 }
1021
1022 /* convert fp to int 't' type */
1023 ST_FUNC void gen_cvt_ftoi(int t)
1024 {
1025 #if 1
1026     gv(RC_FLOAT);
1027     save_reg(TREG_EAX);
1028     save_reg(TREG_EDX);
1029     gen_static_call(TOK___tcc_cvt_ftol);
1030     vtop->r = TREG_EAX; /* mark reg as used */
1031     if (t == VT_LLONG)
1032         vtop->r2 = TREG_EDX;
1033 #else
1034     int bt = vtop->type.t & VT_BTYPE;
1035     if (bt == VT_FLOAT)
1036         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1037     else if (bt == VT_LDOUBLE)
1038         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1039     else
1040         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1041     vswap();
1042     gfunc_call(1);
1043     vpushi(0);
1044     vtop->r = REG_IRET;
1045     vtop->r2 = REG_LRET;
1046 #endif
1047 }
1048
1049 /* convert from one floating point type to another */
1050 ST_FUNC void gen_cvt_ftof(int t)
1051 {
1052     /* all we have to do on i386 is to put the float in a register */
1053     gv(RC_FLOAT);
1054 }
1055
1056 /* computed goto support */
1057 ST_FUNC void ggoto(void)
1058 {
1059     gcall_or_jmp(1);
1060     vtop--;
1061 }
1062
1063 /* bound check support functions */
1064 #ifdef CONFIG_TCC_BCHECK
1065
1066 /* generate a bounded pointer addition */
1067 ST_FUNC void gen_bounded_ptr_add(void)
1068 {
1069     /* prepare fast i386 function call (args in eax and edx) */
1070     gv2(RC_EAX, RC_EDX);
1071     /* save all temporary registers */
1072     vtop -= 2;
1073     save_regs(0);
1074     /* do a fast function call */
1075     gen_static_call(TOK___bound_ptr_add);
1076     /* returned pointer is in eax */
1077     vtop++;
1078     vtop->r = TREG_EAX | VT_BOUNDED;
1079     /* address of bounding function call point */
1080     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1081 }
1082
1083 /* patch pointer addition in vtop so that pointer dereferencing is
1084    also tested */
1085 ST_FUNC void gen_bounded_ptr_deref(void)
1086 {
1087     addr_t func;
1088     int  size, align;
1089     Elf32_Rel *rel;
1090     Sym *sym;
1091
1092     size = 0;
1093     /* XXX: put that code in generic part of tcc */
1094     if (!is_float(vtop->type.t)) {
1095         if (vtop->r & VT_LVAL_BYTE)
1096             size = 1;
1097         else if (vtop->r & VT_LVAL_SHORT)
1098             size = 2;
1099     }
1100     if (!size)
1101         size = type_size(&vtop->type, &align);
1102     switch(size) {
1103     case  1: func = TOK___bound_ptr_indir1; break;
1104     case  2: func = TOK___bound_ptr_indir2; break;
1105     case  4: func = TOK___bound_ptr_indir4; break;
1106     case  8: func = TOK___bound_ptr_indir8; break;
1107     case 12: func = TOK___bound_ptr_indir12; break;
1108     case 16: func = TOK___bound_ptr_indir16; break;
1109     default:
1110         tcc_error("unhandled size when dereferencing bounded pointer");
1111         func = 0;
1112         break;
1113     }
1114
1115     /* patch relocation */
1116     /* XXX: find a better solution ? */
1117     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1118     sym = external_global_sym(func, &func_old_type, 0);
1119     if (!sym->c)
1120         put_extern_sym(sym, NULL, 0, 0);
1121     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1122 }
1123 #endif
1124
1125 /* Save the stack pointer onto the stack */
1126 ST_FUNC void gen_vla_sp_save(int addr) {
1127     /* mov %esp,addr(%ebp)*/
1128     o(0x89);
1129     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1130 }
1131
1132 /* Restore the SP from a location on the stack */
1133 ST_FUNC void gen_vla_sp_restore(int addr) {
1134     o(0x8b);
1135     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1136 }
1137
1138 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1139 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1140 #ifdef TCC_TARGET_PE
1141     /* alloca does more than just adjust %rsp on Windows */
1142     vpush_global_sym(&func_old_type, TOK_alloca);
1143     vswap(); /* Move alloca ref past allocation size */
1144     gfunc_call(1);
1145 #else
1146     int r;
1147     r = gv(RC_INT); /* allocation size */
1148     /* sub r,%rsp */
1149     o(0x2b);
1150     o(0xe0 | r);
1151     /* We align to 16 bytes rather than align */
1152     /* and ~15, %esp */
1153     o(0xf0e483);
1154     vpop();
1155 #endif
1156 }
1157
1158 /* end of X86 code generator */
1159 /*************************************************************/
1160 #endif
1161 /*************************************************************/