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 /* generate jmp to a label */
  74 #define gjmp2(instr,lbl) oad(instr,lbl)
  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 0
  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     if (nocode_wanted)
 104         return;
 105     ind1 = ind + 1;
 106     if (ind1 > cur_text_section->data_allocated)
 107         section_realloc(cur_text_section, ind1);
 108     cur_text_section->data[ind] = c;
 109     ind = ind1;
 110 }
 111
 112 ST_FUNC void o(unsigned int c)
 113 {
 114     while (c) {
 115         g(c);
 116         c = c >> 8;
 117     }
 118 }
 119
 120 ST_FUNC void gen_le16(int v)
 121 {
 122     g(v);
 123     g(v >> 8);
 124 }
 125
 126 ST_FUNC void gen_le32(int c)
 127 {
 128     g(c);
 129     g(c >> 8);
 130     g(c >> 16);
 131     g(c >> 24);
 132 }
 133
 134 /* output a symbol and patch all calls to it */
 135 ST_FUNC void gsym_addr(int t, int a)
 136 {
 137     while (t) {
 138         unsigned char *ptr = cur_text_section->data + t;
 139         uint32_t n = read32le(ptr); /* next value */
 140         write32le(ptr, a - t - 4);
 141         t = n;
 142     }
 143 }
 144
 145 ST_FUNC void gsym(int t)
 146 {
 147     gsym_addr(t, ind);
 148 }
 149
 150 /* instruction + 4 bytes data. Return the address of the data */
 151 ST_FUNC 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 /* output constant with relocation if 'r & VT_SYM' is true */
 163 ST_FUNC void gen_addr32(int r, Sym *sym, long c)
 164 {
 165     if (r & VT_SYM)
 166         greloc(cur_text_section, sym, ind, R_386_32);
 167     gen_le32(c);
 168 }
 169
 170 ST_FUNC void gen_addrpc32(int r, Sym *sym, long c)
 171 {
 172     if (r & VT_SYM)
 173         greloc(cur_text_section, sym, ind, R_386_PC32);
 174     gen_le32(c - 4);
 175 }
 176
 177 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 178    opcode bits */
 179 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 180 {
 181     op_reg = op_reg << 3;
 182     if ((r & VT_VALMASK) == VT_CONST) {
 183         /* constant memory reference */
 184         o(0x05 | op_reg);
 185         gen_addr32(r, sym, c);
 186     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 187         /* currently, we use only ebp as base */
 188         if (c == (char)c) {
 189             /* short reference */
 190             o(0x45 | op_reg);
 191             g(c);
 192         } else {
 193             oad(0x85 | op_reg, c);
 194         }
 195     } else {
 196         g(0x00 | op_reg | (r & VT_VALMASK));
 197     }
 198 }
 199
 200 /* load 'r' from value 'sv' */
 201 ST_FUNC void load(int r, SValue *sv)
 202 {
 203     int v, t, ft, fc, fr;
 204     SValue v1;
 205
 206 #ifdef TCC_TARGET_PE
 207     SValue v2;
 208     sv = pe_getimport(sv, &v2);
 209 #endif
 210
 211     fr = sv->r;
 212     ft = sv->type.t;
 213     fc = sv->c.i;
 214
 215     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 216
 217     v = fr & VT_VALMASK;
 218     if (fr & VT_LVAL) {
 219         if (v == VT_LLOCAL) {
 220             v1.type.t = VT_INT;
 221             v1.r = VT_LOCAL | VT_LVAL;
 222             v1.c.i = fc;
 223             fr = r;
 224             if (!(reg_classes[fr] & RC_INT))
 225                 fr = get_reg(RC_INT);
 226             load(fr, &v1);
 227         }
 228         if ((ft & VT_BTYPE) == VT_FLOAT) {
 229             o(0xd9); /* flds */
 230             r = 0;
 231         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 232             o(0xdd); /* fldl */
 233             r = 0;
 234         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 235             o(0xdb); /* fldt */
 236             r = 5;
 237         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 238             o(0xbe0f);   /* movsbl */
 239         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 240             o(0xb60f);   /* movzbl */
 241         } else if ((ft & VT_TYPE) == VT_SHORT) {
 242             o(0xbf0f);   /* movswl */
 243         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 244             o(0xb70f);   /* movzwl */
 245         } else {
 246             o(0x8b);     /* movl */
 247         }
 248         gen_modrm(r, fr, sv->sym, fc);
 249     } else {
 250         if (v == VT_CONST) {
 251             o(0xb8 + r); /* mov $xx, r */
 252             gen_addr32(fr, sv->sym, fc);
 253         } else if (v == VT_LOCAL) {
 254             if (fc) {
 255                 o(0x8d); /* lea xxx(%ebp), r */
 256                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 257             } else {
 258                 o(0x89);
 259                 o(0xe8 + r); /* mov %ebp, r */
 260             }
 261         } else if (v == VT_CMP) {
 262             oad(0xb8 + r, 0); /* mov $0, r */
 263             o(0x0f); /* setxx %br */
 264             o(fc);
 265             o(0xc0 + r);
 266         } else if (v == VT_JMP || v == VT_JMPI) {
 267             t = v & 1;
 268             oad(0xb8 + r, t); /* mov $1, r */
 269             o(0x05eb); /* jmp after */
 270             gsym(fc);
 271             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 272         } else if (v != r) {
 273             o(0x89);
 274             o(0xc0 + r + v * 8); /* mov v, r */
 275         }
 276     }
 277 }
 278
 279 /* store register 'r' in lvalue 'v' */
 280 ST_FUNC void store(int r, SValue *v)
 281 {
 282     int fr, bt, ft, fc;
 283
 284 #ifdef TCC_TARGET_PE
 285     SValue v2;
 286     v = pe_getimport(v, &v2);
 287 #endif
 288
 289     ft = v->type.t;
 290     fc = v->c.i;
 291     fr = v->r & VT_VALMASK;
 292     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 293     bt = ft & VT_BTYPE;
 294     /* XXX: incorrect if float reg to reg */
 295     if (bt == VT_FLOAT) {
 296         o(0xd9); /* fsts */
 297         r = 2;
 298     } else if (bt == VT_DOUBLE) {
 299         o(0xdd); /* fstpl */
 300         r = 2;
 301     } else if (bt == VT_LDOUBLE) {
 302         o(0xc0d9); /* fld %st(0) */
 303         o(0xdb); /* fstpt */
 304         r = 7;
 305     } else {
 306         if (bt == VT_SHORT)
 307             o(0x66);
 308         if (bt == VT_BYTE || bt == VT_BOOL)
 309             o(0x88);
 310         else
 311             o(0x89);
 312     }
 313     if (fr == VT_CONST ||
 314         fr == VT_LOCAL ||
 315         (v->r & VT_LVAL)) {
 316         gen_modrm(r, v->r, v->sym, fc);
 317     } else if (fr != r) {
 318         o(0xc0 + fr + r * 8); /* mov r, fr */
 319     }
 320 }
 321
 322 static void gadd_sp(int val)
 323 {
 324     if (val == (char)val) {
 325         o(0xc483);
 326         g(val);
 327     } else {
 328         oad(0xc481, val); /* add $xxx, %esp */
 329     }
 330 }
 331
 332 static void gen_static_call(int v)
 333 {
 334     Sym *sym;
 335
 336     sym = external_global_sym(v, &func_old_type, 0);
 337     oad(0xe8, -4);
 338     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 339 }
 340
 341 /* 'is_jmp' is '1' if it is a jump */
 342 static void gcall_or_jmp(int is_jmp)
 343 {
 344     int r;
 345     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 346         int rt;
 347         /* constant case */
 348         if (vtop->r & VT_SYM) {
 349             /* relocation case */
 350             greloc(cur_text_section, vtop->sym,
 351                    ind + 1, R_386_PC32);
 352         } else {
 353             /* put an empty PC32 relocation */
 354             put_elf_reloc(symtab_section, cur_text_section,
 355                           ind + 1, R_386_PC32, 0);
 356         }
 357         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 358         /* extend the return value to the whole register if necessary
 359            visual studio and gcc do not always set the whole eax register
 360            when assigning the return value of a function  */
 361         rt = vtop->type.ref->type.t;
 362         switch (rt & VT_BTYPE) {
 363             case VT_BYTE:
 364                 if (rt & VT_UNSIGNED) {
 365                     o(0xc0b60f); /* movzx %al, %eax */
 366                 }
 367                 else {
 368                     o(0xc0be0f); /* movsx %al, %eax */
 369                 }
 370                 break;
 371             case VT_SHORT:
 372                 if (rt & VT_UNSIGNED) {
 373                     o(0xc0b70f); /* movzx %ax, %eax */
 374                 }
 375                 else {
 376                     o(0xc0bf0f); /* movsx %ax, %eax */
 377                 }
 378                 break;
 379             default:
 380                 break;
 381         }
 382     } else {
 383         /* otherwise, indirect call */
 384         r = gv(RC_INT);
 385         o(0xff); /* call/jmp *r */
 386         o(0xd0 + r + (is_jmp << 4));
 387     }
 388 }
 389
 390 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 391 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 392
 393 /* Return the number of registers needed to return the struct, or 0 if
 394    returning via struct pointer. */
 395 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 396 {
 397 #ifdef TCC_TARGET_PE
 398     int size, align;
 399     *ret_align = 1; // Never have to re-align return values for x86
 400     *regsize = 4;
 401     size = type_size(vt, &align);
 402     if (size > 8 || (size & (size - 1)))
 403         return 0;
 404     if (size == 8)
 405         ret->t = VT_LLONG;
 406     else if (size == 4)
 407         ret->t = VT_INT;
 408     else if (size == 2)
 409         ret->t = VT_SHORT;
 410     else
 411         ret->t = VT_BYTE;
 412     ret->ref = NULL;
 413     return 1;
 414 #else
 415     *ret_align = 1; // Never have to re-align return values for x86
 416     return 0;
 417 #endif
 418 }
 419
 420 /* Generate function call. The function address is pushed first, then
 421    all the parameters in call order. This functions pops all the
 422    parameters and the function address. */
 423 ST_FUNC void gfunc_call(int nb_args)
 424 {
 425     int size, align, r, args_size, i, func_call;
 426     Sym *func_sym;
 427
 428     args_size = 0;
 429     for(i = 0;i < nb_args; i++) {
 430         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 431             size = type_size(&vtop->type, &align);
 432             /* align to stack align size */
 433             size = (size + 3) & ~3;
 434             /* allocate the necessary size on stack */
 435             oad(0xec81, size); /* sub $xxx, %esp */
 436             /* generate structure store */
 437             r = get_reg(RC_INT);
 438             o(0x89); /* mov %esp, r */
 439             o(0xe0 + r);
 440             vset(&vtop->type, r | VT_LVAL, 0);
 441             vswap();
 442             vstore();
 443             args_size += size;
 444         } else if (is_float(vtop->type.t)) {
 445             gv(RC_FLOAT); /* only one float register */
 446             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 447                 size = 4;
 448             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 449                 size = 8;
 450             else
 451                 size = 12;
 452             oad(0xec81, size); /* sub $xxx, %esp */
 453             if (size == 12)
 454                 o(0x7cdb);
 455             else
 456                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 457             g(0x24);
 458             g(0x00);
 459             args_size += size;
 460         } else {
 461             /* simple type (currently always same size) */
 462             /* XXX: implicit cast ? */
 463             r = gv(RC_INT);
 464             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 465                 size = 8;
 466                 o(0x50 + vtop->r2); /* push r */
 467             } else {
 468                 size = 4;
 469             }
 470             o(0x50 + r); /* push r */
 471             args_size += size;
 472         }
 473         vtop--;
 474     }
 475     save_regs(0); /* save used temporary registers */
 476     func_sym = vtop->type.ref;
 477     func_call = func_sym->a.func_call;
 478     /* fast call case */
 479     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 480         func_call == FUNC_FASTCALLW) {
 481         int fastcall_nb_regs;
 482         uint8_t *fastcall_regs_ptr;
 483         if (func_call == FUNC_FASTCALLW) {
 484             fastcall_regs_ptr = fastcallw_regs;
 485             fastcall_nb_regs = 2;
 486         } else {
 487             fastcall_regs_ptr = fastcall_regs;
 488             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 489         }
 490         for(i = 0;i < fastcall_nb_regs; i++) {
 491             if (args_size <= 0)
 492                 break;
 493             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 494             /* XXX: incorrect for struct/floats */
 495             args_size -= 4;
 496         }
 497     }
 498 #ifndef TCC_TARGET_PE
 499     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 500         args_size -= 4;
 501 #endif
 502     gcall_or_jmp(0);
 503
 504     if (args_size && func_call != FUNC_STDCALL)
 505         gadd_sp(args_size);
 506     vtop--;
 507 }
 508
 509 #ifdef TCC_TARGET_PE
 510 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 511 #else
 512 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 513 #endif
 514
 515 /* generate function prolog of type 't' */
 516 ST_FUNC void gfunc_prolog(CType *func_type)
 517 {
 518     int addr, align, size, func_call, fastcall_nb_regs;
 519     int param_index, param_addr;
 520     uint8_t *fastcall_regs_ptr;
 521     Sym *sym;
 522     CType *type;
 523
 524     sym = func_type->ref;
 525     func_call = sym->a.func_call;
 526     addr = 8;
 527     loc = 0;
 528     func_vc = 0;
 529
 530     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 531         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 532         fastcall_regs_ptr = fastcall_regs;
 533     } else if (func_call == FUNC_FASTCALLW) {
 534         fastcall_nb_regs = 2;
 535         fastcall_regs_ptr = fastcallw_regs;
 536     } else {
 537         fastcall_nb_regs = 0;
 538         fastcall_regs_ptr = NULL;
 539     }
 540     param_index = 0;
 541
 542     ind += FUNC_PROLOG_SIZE;
 543     func_sub_sp_offset = ind;
 544     /* if the function returns a structure, then add an
 545        implicit pointer parameter */
 546     func_vt = sym->type;
 547     func_var = (sym->c == FUNC_ELLIPSIS);
 548 #ifdef TCC_TARGET_PE
 549     size = type_size(&func_vt,&align);
 550     if (((func_vt.t & VT_BTYPE) == VT_STRUCT)
 551         && (size > 8 || (size & (size - 1)))) {
 552 #else
 553     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 554 #endif
 555         /* XXX: fastcall case ? */
 556         func_vc = addr;
 557         addr += 4;
 558         param_index++;
 559     }
 560     /* define parameters */
 561     while ((sym = sym->next) != NULL) {
 562         type = &sym->type;
 563         size = type_size(type, &align);
 564         size = (size + 3) & ~3;
 565 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 566         /* structs are passed as pointer */
 567         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 568             size = 4;
 569         }
 570 #endif
 571         if (param_index < fastcall_nb_regs) {
 572             /* save FASTCALL register */
 573             loc -= 4;
 574             o(0x89);     /* movl */
 575             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 576             param_addr = loc;
 577         } else {
 578             param_addr = addr;
 579             addr += size;
 580         }
 581         sym_push(sym->v & ~SYM_FIELD, type,
 582                  VT_LOCAL | lvalue_type(type->t), param_addr);
 583         param_index++;
 584     }
 585     func_ret_sub = 0;
 586     /* pascal type call ? */
 587     if (func_call == FUNC_STDCALL)
 588         func_ret_sub = addr - 8;
 589 #ifndef TCC_TARGET_PE
 590     else if (func_vc)
 591         func_ret_sub = 4;
 592 #endif
 593
 594 #ifdef CONFIG_TCC_BCHECK
 595     /* leave some room for bound checking code */
 596     if (tcc_state->do_bounds_check) {
 597         func_bound_offset = lbounds_section->data_offset;
 598         func_bound_ind = ind;
 599         oad(0xb8, 0); /* lbound section pointer */
 600         oad(0xb8, 0); /* call to function */
 601     }
 602 #endif
 603 }
 604
 605 /* generate function epilog */
 606 ST_FUNC void gfunc_epilog(void)
 607 {
 608     addr_t v, saved_ind;
 609
 610 #ifdef CONFIG_TCC_BCHECK
 611     if (tcc_state->do_bounds_check
 612      && func_bound_offset != lbounds_section->data_offset) {
 613         addr_t saved_ind;
 614         addr_t *bounds_ptr;
 615         Sym *sym_data;
 616
 617         /* add end of table info */
 618         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 619         *bounds_ptr = 0;
 620
 621         /* generate bound local allocation */
 622         saved_ind = ind;
 623         ind = func_bound_ind;
 624         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 625                                func_bound_offset, lbounds_section->data_offset);
 626         greloc(cur_text_section, sym_data,
 627                ind + 1, R_386_32);
 628         oad(0xb8, 0); /* mov %eax, xxx */
 629         gen_static_call(TOK___bound_local_new);
 630         ind = saved_ind;
 631
 632         /* generate bound check local freeing */
 633         o(0x5250); /* save returned value, if any */
 634         greloc(cur_text_section, sym_data, ind + 1, R_386_32);
 635         oad(0xb8, 0); /* mov %eax, xxx */
 636         gen_static_call(TOK___bound_local_delete);
 637         o(0x585a); /* restore returned value, if any */
 638     }
 639 #endif
 640
 641     /* align local size to word & save local variables */
 642     v = (-loc + 3) & -4;
 643
 644 #if USE_EBX
 645     o(0x8b);
 646     gen_modrm(TREG_EBX, VT_LOCAL, NULL, -(v+4));
 647 #endif
 648
 649     o(0xc9); /* leave */
 650     if (func_ret_sub == 0) {
 651         o(0xc3); /* ret */
 652     } else {
 653         o(0xc2); /* ret n */
 654         g(func_ret_sub);
 655         g(func_ret_sub >> 8);
 656     }
 657     saved_ind = ind;
 658     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 659 #ifdef TCC_TARGET_PE
 660     if (v >= 4096) {
 661         oad(0xb8, v); /* mov stacksize, %eax */
 662         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 663     } else
 664 #endif
 665     {
 666         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 667         o(0xec81);  /* sub esp, stacksize */
 668         gen_le32(v);
 669 #ifdef TCC_TARGET_PE
 670         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 671 #endif
 672     }
 673     o(0x53 * USE_EBX); /* push ebx */
 674     ind = saved_ind;
 675 }
 676
 677 /* generate a jump to a label */
 678 ST_FUNC int gjmp(int t)
 679 {
 680     return gjmp2(0xe9, t);
 681 }
 682
 683 /* generate a jump to a fixed address */
 684 ST_FUNC void gjmp_addr(int a)
 685 {
 686     int r;
 687     r = a - ind - 2;
 688     if (r == (char)r) {
 689         g(0xeb);
 690         g(r);
 691     } else {
 692         oad(0xe9, a - ind - 5);
 693     }
 694 }
 695
 696 ST_FUNC void gtst_addr(int inv, int a)
 697 {
 698     int v = vtop->r & VT_VALMASK;
 699     if (v == VT_CMP) {
 700         inv ^= (vtop--)->c.i;
 701         a -= ind + 2;
 702         if (a == (char)a) {
 703             g(inv - 32);
 704             g(a);
 705         } else {
 706             g(0x0f);
 707             oad(inv - 16, a - 4);
 708         }
 709     } else if ((v & ~1) == VT_JMP) {
 710         if ((v & 1) != inv) {
 711             gjmp_addr(a);
 712             gsym(vtop->c.i);
 713         } else {
 714             gsym(vtop->c.i);
 715             o(0x05eb);
 716             gjmp_addr(a);
 717         }
 718         vtop--;
 719     }
 720 }
 721
 722 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 723 ST_FUNC int gtst(int inv, int t)
 724 {
 725     int v = vtop->r & VT_VALMASK;
 726     if (nocode_wanted) {
 727         ;
 728     } else if (v == VT_CMP) {
 729         /* fast case : can jump directly since flags are set */
 730         g(0x0f);
 731         t = gjmp2((vtop->c.i - 16) ^ inv, t);
 732     } else if (v == VT_JMP || v == VT_JMPI) {
 733         /* && or || optimization */
 734         if ((v & 1) == inv) {
 735             /* insert vtop->c jump list in t */
 736             uint32_t n1, n = vtop->c.i;
 737             if (n) {
 738                 while ((n1 = read32le(cur_text_section->data + n)))
 739                     n = n1;
 740                 write32le(cur_text_section->data + n, t);
 741                 t = vtop->c.i;
 742             }
 743         } else {
 744             t = gjmp(t);
 745             gsym(vtop->c.i);
 746         }
 747     }
 748     vtop--;
 749     return t;
 750 }
 751
 752 /* generate an integer binary operation */
 753 ST_FUNC void gen_opi(int op)
 754 {
 755     int r, fr, opc, c;
 756
 757     switch(op) {
 758     case '+':
 759     case TOK_ADDC1: /* add with carry generation */
 760         opc = 0;
 761     gen_op8:
 762         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 763             /* constant case */
 764             vswap();
 765             r = gv(RC_INT);
 766             vswap();
 767             c = vtop->c.i;
 768             if (c == (char)c) {
 769                 /* generate inc and dec for smaller code */
 770                 if (c==1 && opc==0 && op != TOK_ADDC1) {
 771                     o (0x40 | r); // inc
 772                 } else if (c==1 && opc==5 && op != TOK_SUBC1) {
 773                     o (0x48 | r); // dec
 774                 } else {
 775                     o(0x83);
 776                     o(0xc0 | (opc << 3) | r);
 777                     g(c);
 778                 }
 779             } else {
 780                 o(0x81);
 781                 oad(0xc0 | (opc << 3) | r, c);
 782             }
 783         } else {
 784             gv2(RC_INT, RC_INT);
 785             r = vtop[-1].r;
 786             fr = vtop[0].r;
 787             o((opc << 3) | 0x01);
 788             o(0xc0 + r + fr * 8);
 789         }
 790         vtop--;
 791         if (op >= TOK_ULT && op <= TOK_GT) {
 792             vtop->r = VT_CMP;
 793             vtop->c.i = op;
 794         }
 795         break;
 796     case '-':
 797     case TOK_SUBC1: /* sub with carry generation */
 798         opc = 5;
 799         goto gen_op8;
 800     case TOK_ADDC2: /* add with carry use */
 801         opc = 2;
 802         goto gen_op8;
 803     case TOK_SUBC2: /* sub with carry use */
 804         opc = 3;
 805         goto gen_op8;
 806     case '&':
 807         opc = 4;
 808         goto gen_op8;
 809     case '^':
 810         opc = 6;
 811         goto gen_op8;
 812     case '|':
 813         opc = 1;
 814         goto gen_op8;
 815     case '*':
 816         gv2(RC_INT, RC_INT);
 817         r = vtop[-1].r;
 818         fr = vtop[0].r;
 819         vtop--;
 820         o(0xaf0f); /* imul fr, r */
 821         o(0xc0 + fr + r * 8);
 822         break;
 823     case TOK_SHL:
 824         opc = 4;
 825         goto gen_shift;
 826     case TOK_SHR:
 827         opc = 5;
 828         goto gen_shift;
 829     case TOK_SAR:
 830         opc = 7;
 831     gen_shift:
 832         opc = 0xc0 | (opc << 3);
 833         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 834             /* constant case */
 835             vswap();
 836             r = gv(RC_INT);
 837             vswap();
 838             c = vtop->c.i & 0x1f;
 839             o(0xc1); /* shl/shr/sar $xxx, r */
 840             o(opc | r);
 841             g(c);
 842         } else {
 843             /* we generate the shift in ecx */
 844             gv2(RC_INT, RC_ECX);
 845             r = vtop[-1].r;
 846             o(0xd3); /* shl/shr/sar %cl, r */
 847             o(opc | r);
 848         }
 849         vtop--;
 850         break;
 851     case '/':
 852     case TOK_UDIV:
 853     case TOK_PDIV:
 854     case '%':
 855     case TOK_UMOD:
 856     case TOK_UMULL:
 857         /* first operand must be in eax */
 858         /* XXX: need better constraint for second operand */
 859         gv2(RC_EAX, RC_ECX);
 860         r = vtop[-1].r;
 861         fr = vtop[0].r;
 862         vtop--;
 863         save_reg(TREG_EDX);
 864         /* save EAX too if used otherwise */
 865         save_reg_upstack(TREG_EAX, 1);
 866         if (op == TOK_UMULL) {
 867             o(0xf7); /* mul fr */
 868             o(0xe0 + fr);
 869             vtop->r2 = TREG_EDX;
 870             r = TREG_EAX;
 871         } else {
 872             if (op == TOK_UDIV || op == TOK_UMOD) {
 873                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 874                 o(0xf0 + fr);
 875             } else {
 876                 o(0xf799); /* cltd, idiv fr, %eax */
 877                 o(0xf8 + fr);
 878             }
 879             if (op == '%' || op == TOK_UMOD)
 880                 r = TREG_EDX;
 881             else
 882                 r = TREG_EAX;
 883         }
 884         vtop->r = r;
 885         break;
 886     default:
 887         opc = 7;
 888         goto gen_op8;
 889     }
 890 }
 891
 892 /* generate a floating point operation 'v = t1 op t2' instruction. The
 893    two operands are guaranted to have the same floating point type */
 894 /* XXX: need to use ST1 too */
 895 ST_FUNC void gen_opf(int op)
 896 {
 897     int a, ft, fc, swapped, r;
 898
 899     /* convert constants to memory references */
 900     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 901         vswap();
 902         gv(RC_FLOAT);
 903         vswap();
 904     }
 905     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 906         gv(RC_FLOAT);
 907
 908     /* must put at least one value in the floating point register */
 909     if ((vtop[-1].r & VT_LVAL) &&
 910         (vtop[0].r & VT_LVAL)) {
 911         vswap();
 912         gv(RC_FLOAT);
 913         vswap();
 914     }
 915     swapped = 0;
 916     /* swap the stack if needed so that t1 is the register and t2 is
 917        the memory reference */
 918     if (vtop[-1].r & VT_LVAL) {
 919         vswap();
 920         swapped = 1;
 921     }
 922     if (op >= TOK_ULT && op <= TOK_GT) {
 923         /* load on stack second operand */
 924         load(TREG_ST0, vtop);
 925         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 926         if (op == TOK_GE || op == TOK_GT)
 927             swapped = !swapped;
 928         else if (op == TOK_EQ || op == TOK_NE)
 929             swapped = 0;
 930         if (swapped)
 931             o(0xc9d9); /* fxch %st(1) */
 932         if (op == TOK_EQ || op == TOK_NE)
 933             o(0xe9da); /* fucompp */
 934         else
 935             o(0xd9de); /* fcompp */
 936         o(0xe0df); /* fnstsw %ax */
 937         if (op == TOK_EQ) {
 938             o(0x45e480); /* and $0x45, %ah */
 939             o(0x40fC80); /* cmp $0x40, %ah */
 940         } else if (op == TOK_NE) {
 941             o(0x45e480); /* and $0x45, %ah */
 942             o(0x40f480); /* xor $0x40, %ah */
 943             op = TOK_NE;
 944         } else if (op == TOK_GE || op == TOK_LE) {
 945             o(0x05c4f6); /* test $0x05, %ah */
 946             op = TOK_EQ;
 947         } else {
 948             o(0x45c4f6); /* test $0x45, %ah */
 949             op = TOK_EQ;
 950         }
 951         vtop--;
 952         vtop->r = VT_CMP;
 953         vtop->c.i = op;
 954     } else {
 955         /* no memory reference possible for long double operations */
 956         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 957             load(TREG_ST0, vtop);
 958             swapped = !swapped;
 959         }
 960
 961         switch(op) {
 962         default:
 963         case '+':
 964             a = 0;
 965             break;
 966         case '-':
 967             a = 4;
 968             if (swapped)
 969                 a++;
 970             break;
 971         case '*':
 972             a = 1;
 973             break;
 974         case '/':
 975             a = 6;
 976             if (swapped)
 977                 a++;
 978             break;
 979         }
 980         ft = vtop->type.t;
 981         fc = vtop->c.i;
 982         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 983             o(0xde); /* fxxxp %st, %st(1) */
 984             o(0xc1 + (a << 3));
 985         } else {
 986             /* if saved lvalue, then we must reload it */
 987             r = vtop->r;
 988             if ((r & VT_VALMASK) == VT_LLOCAL) {
 989                 SValue v1;
 990                 r = get_reg(RC_INT);
 991                 v1.type.t = VT_INT;
 992                 v1.r = VT_LOCAL | VT_LVAL;
 993                 v1.c.i = fc;
 994                 load(r, &v1);
 995                 fc = 0;
 996             }
 997
 998             if ((ft & VT_BTYPE) == VT_DOUBLE)
 999                 o(0xdc);
1000             else
1001                 o(0xd8);
1002             gen_modrm(a, r, vtop->sym, fc);
1003         }
1004         vtop--;
1005     }
1006 }
1007
1008 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1009    and 'long long' cases. */
1010 ST_FUNC void gen_cvt_itof(int t)
1011 {
1012     save_reg(TREG_ST0);
1013     gv(RC_INT);
1014     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
1015         /* signed long long to float/double/long double (unsigned case
1016            is handled generically) */
1017         o(0x50 + vtop->r2); /* push r2 */
1018         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1019         o(0x242cdf); /* fildll (%esp) */
1020         o(0x08c483); /* add $8, %esp */
1021     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
1022                (VT_INT | VT_UNSIGNED)) {
1023         /* unsigned int to float/double/long double */
1024         o(0x6a); /* push $0 */
1025         g(0x00);
1026         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1027         o(0x242cdf); /* fildll (%esp) */
1028         o(0x08c483); /* add $8, %esp */
1029     } else {
1030         /* int to float/double/long double */
1031         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1032         o(0x2404db); /* fildl (%esp) */
1033         o(0x04c483); /* add $4, %esp */
1034     }
1035     vtop->r = TREG_ST0;
1036 }
1037
1038 /* convert fp to int 't' type */
1039 ST_FUNC void gen_cvt_ftoi(int t)
1040 {
1041     int bt = vtop->type.t & VT_BTYPE;
1042     if (bt == VT_FLOAT)
1043         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1044     else if (bt == VT_LDOUBLE)
1045         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1046     else
1047         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1048     vswap();
1049     gfunc_call(1);
1050     vpushi(0);
1051     vtop->r = REG_IRET;
1052     vtop->r2 = REG_LRET;
1053 }
1054
1055 /* convert from one floating point type to another */
1056 ST_FUNC void gen_cvt_ftof(int t)
1057 {
1058     /* all we have to do on i386 is to put the float in a register */
1059     gv(RC_FLOAT);
1060 }
1061
1062 /* computed goto support */
1063 ST_FUNC void ggoto(void)
1064 {
1065     gcall_or_jmp(1);
1066     vtop--;
1067 }
1068
1069 /* bound check support functions */
1070 #ifdef CONFIG_TCC_BCHECK
1071
1072 /* generate a bounded pointer addition */
1073 ST_FUNC void gen_bounded_ptr_add(void)
1074 {
1075     /* prepare fast i386 function call (args in eax and edx) */
1076     gv2(RC_EAX, RC_EDX);
1077     /* save all temporary registers */
1078     vtop -= 2;
1079     save_regs(0);
1080     /* do a fast function call */
1081     gen_static_call(TOK___bound_ptr_add);
1082     /* returned pointer is in eax */
1083     vtop++;
1084     vtop->r = TREG_EAX | VT_BOUNDED;
1085     /* address of bounding function call point */
1086     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1087 }
1088
1089 /* patch pointer addition in vtop so that pointer dereferencing is
1090    also tested */
1091 ST_FUNC void gen_bounded_ptr_deref(void)
1092 {
1093     addr_t func;
1094     int  size, align;
1095     Elf32_Rel *rel;
1096     Sym *sym;
1097
1098     size = 0;
1099     /* XXX: put that code in generic part of tcc */
1100     if (!is_float(vtop->type.t)) {
1101         if (vtop->r & VT_LVAL_BYTE)
1102             size = 1;
1103         else if (vtop->r & VT_LVAL_SHORT)
1104             size = 2;
1105     }
1106     if (!size)
1107         size = type_size(&vtop->type, &align);
1108     switch(size) {
1109     case  1: func = TOK___bound_ptr_indir1; break;
1110     case  2: func = TOK___bound_ptr_indir2; break;
1111     case  4: func = TOK___bound_ptr_indir4; break;
1112     case  8: func = TOK___bound_ptr_indir8; break;
1113     case 12: func = TOK___bound_ptr_indir12; break;
1114     case 16: func = TOK___bound_ptr_indir16; break;
1115     default:
1116         tcc_error("unhandled size when dereferencing bounded pointer");
1117         func = 0;
1118         break;
1119     }
1120
1121     /* patch relocation */
1122     /* XXX: find a better solution ? */
1123     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1124     sym = external_global_sym(func, &func_old_type, 0);
1125     if (!sym->c)
1126         put_extern_sym(sym, NULL, 0, 0);
1127     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1128 }
1129 #endif
1130
1131 /* Save the stack pointer onto the stack */
1132 ST_FUNC void gen_vla_sp_save(int addr) {
1133     /* mov %esp,addr(%ebp)*/
1134     o(0x89);
1135     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1136 }
1137
1138 /* Restore the SP from a location on the stack */
1139 ST_FUNC void gen_vla_sp_restore(int addr) {
1140     o(0x8b);
1141     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1142 }
1143
1144 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1145 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1146 #ifdef TCC_TARGET_PE
1147     /* alloca does more than just adjust %rsp on Windows */
1148     vpush_global_sym(&func_old_type, TOK_alloca);
1149     vswap(); /* Move alloca ref past allocation size */
1150     gfunc_call(1);
1151 #else
1152     int r;
1153     r = gv(RC_INT); /* allocation size */
1154     /* sub r,%rsp */
1155     o(0x2b);
1156     o(0xe0 | r);
1157     /* We align to 16 bytes rather than align */
1158     /* and ~15, %esp */
1159     o(0xf0e483);
1160     vpop();
1161 #endif
1162 }
1163
1164 /* end of X86 code generator */
1165 /*************************************************************/
1166 #endif
1167 /*************************************************************/