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 /* ELF defines */
  78
  79 #define EM_TCC_TARGET EM_386
  80
  81 /* relocation type for 32 bit data relocation */
  82 #define R_DATA_32   R_386_32
  83 #define R_DATA_PTR  R_386_32
  84 #define R_JMP_SLOT  R_386_JMP_SLOT
  85 #define R_COPY      R_386_COPY
  86
  87 #define ELF_START_ADDR 0x08048000
  88 #define ELF_PAGE_SIZE  0x1000
  89
  90 /******************************************************/
  91 #else /* ! TARGET_DEFS_ONLY */
  92 /******************************************************/
  93 #include "tcc.h"
  94
  95 /* define to 1/0 to [not] have EBX as 4th register */
  96 #define USE_EBX 1
  97
  98 ST_DATA const int reg_classes[NB_REGS] = {
  99     /* eax */ RC_INT | RC_EAX,
 100     /* ecx */ RC_INT | RC_ECX,
 101     /* edx */ RC_INT | RC_EDX,
 102     /* ebx */ (RC_INT | RC_EBX) * USE_EBX,
 103     /* st0 */ RC_FLOAT | RC_ST0,
 104 };
 105
 106 static unsigned long func_sub_sp_offset;
 107 static int func_ret_sub;
 108 #ifdef CONFIG_TCC_BCHECK
 109 static addr_t func_bound_offset;
 110 static unsigned long func_bound_ind;
 111 #endif
 112
 113 /* XXX: make it faster ? */
 114 ST_FUNC void g(int c)
 115 {
 116     int ind1;
 117     ind1 = ind + 1;
 118     if (ind1 > cur_text_section->data_allocated)
 119         section_realloc(cur_text_section, ind1);
 120     cur_text_section->data[ind] = c;
 121     ind = ind1;
 122 }
 123
 124 ST_FUNC void o(unsigned int c)
 125 {
 126     while (c) {
 127         g(c);
 128         c = c >> 8;
 129     }
 130 }
 131
 132 ST_FUNC void gen_le16(int v)
 133 {
 134     g(v);
 135     g(v >> 8);
 136 }
 137
 138 ST_FUNC void gen_le32(int c)
 139 {
 140     g(c);
 141     g(c >> 8);
 142     g(c >> 16);
 143     g(c >> 24);
 144 }
 145
 146 /* output a symbol and patch all calls to it */
 147 ST_FUNC void gsym_addr(int t, int a)
 148 {
 149     while (t) {
 150         unsigned char *ptr = cur_text_section->data + t;
 151         uint32_t n = read32le(ptr); /* next value */
 152         write32le(ptr, a - t - 4);
 153         t = n;
 154     }
 155 }
 156
 157 ST_FUNC void gsym(int t)
 158 {
 159     gsym_addr(t, ind);
 160 }
 161
 162 /* psym is used to put an instruction with a data field which is a
 163    reference to a symbol. It is in fact the same as oad ! */
 164 #define psym oad
 165
 166 /* instruction + 4 bytes data. Return the address of the data */
 167 ST_FUNC int oad(int c, int s)
 168 {
 169     int ind1;
 170
 171     o(c);
 172     ind1 = ind + 4;
 173     if (ind1 > cur_text_section->data_allocated)
 174         section_realloc(cur_text_section, ind1);
 175     write32le(cur_text_section->data + ind, s);
 176     s = ind;
 177     ind = ind1;
 178     return s;
 179 }
 180
 181 /* output constant with relocation if 'r & VT_SYM' is true */
 182 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 183 {
 184     if (r & VT_SYM)
 185         greloc(cur_text_section, sym, ind, R_386_32);
 186     gen_le32(c);
 187 }
 188
 189 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 190 {
 191     if (r & VT_SYM)
 192         greloc(cur_text_section, sym, ind, R_386_PC32);
 193     gen_le32(c - 4);
 194 }
 195
 196 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 197    opcode bits */
 198 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 199 {
 200     op_reg = op_reg << 3;
 201     if ((r & VT_VALMASK) == VT_CONST) {
 202         /* constant memory reference */
 203         o(0x05 | op_reg);
 204         gen_addr32(r, sym, c);
 205     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 206         /* currently, we use only ebp as base */
 207         if (c == (char)c) {
 208             /* short reference */
 209             o(0x45 | op_reg);
 210             g(c);
 211         } else {
 212             oad(0x85 | op_reg, c);
 213         }
 214     } else {
 215         g(0x00 | op_reg | (r & VT_VALMASK));
 216     }
 217 }
 218
 219 /* load 'r' from value 'sv' */
 220 ST_FUNC void load(int r, SValue *sv)
 221 {
 222     int v, t, ft, fc, fr;
 223     SValue v1;
 224
 225 #ifdef TCC_TARGET_PE
 226     SValue v2;
 227     sv = pe_getimport(sv, &v2);
 228 #endif
 229
 230     fr = sv->r;
 231     ft = sv->type.t;
 232     fc = sv->c.i;
 233
 234     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 235
 236     v = fr & VT_VALMASK;
 237     if (fr & VT_LVAL) {
 238         if (v == VT_LLOCAL) {
 239             v1.type.t = VT_INT;
 240             v1.r = VT_LOCAL | VT_LVAL;
 241             v1.c.i = fc;
 242             fr = r;
 243             if (!(reg_classes[fr] & RC_INT))
 244                 fr = get_reg(RC_INT);
 245             load(fr, &v1);
 246         }
 247         if ((ft & VT_BTYPE) == VT_FLOAT) {
 248             o(0xd9); /* flds */
 249             r = 0;
 250         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 251             o(0xdd); /* fldl */
 252             r = 0;
 253         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 254             o(0xdb); /* fldt */
 255             r = 5;
 256         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 257             o(0xbe0f);   /* movsbl */
 258         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 259             o(0xb60f);   /* movzbl */
 260         } else if ((ft & VT_TYPE) == VT_SHORT) {
 261             o(0xbf0f);   /* movswl */
 262         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 263             o(0xb70f);   /* movzwl */
 264         } else {
 265             o(0x8b);     /* movl */
 266         }
 267         gen_modrm(r, fr, sv->sym, fc);
 268     } else {
 269         if (v == VT_CONST) {
 270             o(0xb8 + r); /* mov $xx, r */
 271             gen_addr32(fr, sv->sym, fc);
 272         } else if (v == VT_LOCAL) {
 273             if (fc) {
 274                 o(0x8d); /* lea xxx(%ebp), r */
 275                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 276             } else {
 277                 o(0x89);
 278                 o(0xe8 + r); /* mov %ebp, r */
 279             }
 280         } else if (v == VT_CMP) {
 281             oad(0xb8 + r, 0); /* mov $0, r */
 282             o(0x0f); /* setxx %br */
 283             o(fc);
 284             o(0xc0 + r);
 285         } else if (v == VT_JMP || v == VT_JMPI) {
 286             t = v & 1;
 287             oad(0xb8 + r, t); /* mov $1, r */
 288             o(0x05eb); /* jmp after */
 289             gsym(fc);
 290             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 291         } else if (v != r) {
 292             o(0x89);
 293             o(0xc0 + r + v * 8); /* mov v, r */
 294         }
 295     }
 296 }
 297
 298 /* store register 'r' in lvalue 'v' */
 299 ST_FUNC void store(int r, SValue *v)
 300 {
 301     int fr, bt, ft, fc;
 302
 303 #ifdef TCC_TARGET_PE
 304     SValue v2;
 305     v = pe_getimport(v, &v2);
 306 #endif
 307
 308     ft = v->type.t;
 309     fc = v->c.i;
 310     fr = v->r & VT_VALMASK;
 311     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 312     bt = ft & VT_BTYPE;
 313     /* XXX: incorrect if float reg to reg */
 314     if (bt == VT_FLOAT) {
 315         o(0xd9); /* fsts */
 316         r = 2;
 317     } else if (bt == VT_DOUBLE) {
 318         o(0xdd); /* fstpl */
 319         r = 2;
 320     } else if (bt == VT_LDOUBLE) {
 321         o(0xc0d9); /* fld %st(0) */
 322         o(0xdb); /* fstpt */
 323         r = 7;
 324     } else {
 325         if (bt == VT_SHORT)
 326             o(0x66);
 327         if (bt == VT_BYTE || bt == VT_BOOL)
 328             o(0x88);
 329         else
 330             o(0x89);
 331     }
 332     if (fr == VT_CONST ||
 333         fr == VT_LOCAL ||
 334         (v->r & VT_LVAL)) {
 335         gen_modrm(r, v->r, v->sym, fc);
 336     } else if (fr != r) {
 337         o(0xc0 + fr + r * 8); /* mov r, fr */
 338     }
 339 }
 340
 341 static void gadd_sp(int val)
 342 {
 343     if (val == (char)val) {
 344         o(0xc483);
 345         g(val);
 346     } else {
 347         oad(0xc481, val); /* add $xxx, %esp */
 348     }
 349 }
 350
 351 static void gen_static_call(int v)
 352 {
 353     Sym *sym;
 354
 355     sym = external_global_sym(v, &func_old_type, 0);
 356     oad(0xe8, -4);
 357     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 358 }
 359
 360 /* 'is_jmp' is '1' if it is a jump */
 361 static void gcall_or_jmp(int is_jmp)
 362 {
 363     int r;
 364     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 365         int rt;
 366         /* constant case */
 367         if (vtop->r & VT_SYM) {
 368             /* relocation case */
 369             greloc(cur_text_section, vtop->sym,
 370                    ind + 1, R_386_PC32);
 371         } else {
 372             /* put an empty PC32 relocation */
 373             put_elf_reloc(symtab_section, cur_text_section,
 374                           ind + 1, R_386_PC32, 0);
 375         }
 376         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 377         /* extend the return value to the whole register if necessary
 378            visual studio and gcc do not always set the whole eax register
 379            when assigning the return value of a function  */
 380         rt = vtop->type.ref->type.t;
 381         switch (rt & VT_BTYPE) {
 382             case VT_BYTE:
 383                 if (rt & VT_UNSIGNED) {
 384                     o(0xc0b60f); /* movzx %al, %eax */
 385                 }
 386                 else {
 387                     o(0xc0be0f); /* movsx %al, %eax */
 388                 }
 389                 break;
 390             case VT_SHORT:
 391                 if (rt & VT_UNSIGNED) {
 392                     o(0xc0b70f); /* movzx %ax, %eax */
 393                 }
 394                 else {
 395                     o(0xc0bf0f); /* movsx %ax, %eax */
 396                 }
 397                 break;
 398             default:
 399                 break;
 400         }
 401     } else {
 402         /* otherwise, indirect call */
 403         r = gv(RC_INT);
 404         o(0xff); /* call/jmp *r */
 405         o(0xd0 + r + (is_jmp << 4));
 406     }
 407 }
 408
 409 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 410 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 411
 412 /* Return the number of registers needed to return the struct, or 0 if
 413    returning via struct pointer. */
 414 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 415 {
 416 #ifdef TCC_TARGET_PE
 417     int size, align;
 418
 419     *ret_align = 1; // Never have to re-align return values for x86
 420     *regsize = 4;
 421     size = type_size(vt, &align);
 422     if (size > 8) {
 423         return 0;
 424     } else if (size > 4) {
 425         ret->ref = NULL;
 426         ret->t = VT_LLONG;
 427         return 1;
 428     } else {
 429         ret->ref = NULL;
 430         ret->t = VT_INT;
 431         return 1;
 432     }
 433 #else
 434     *ret_align = 1; // Never have to re-align return values for x86
 435     return 0;
 436 #endif
 437 }
 438
 439 /* Generate function call. The function address is pushed first, then
 440    all the parameters in call order. This functions pops all the
 441    parameters and the function address. */
 442 ST_FUNC void gfunc_call(int nb_args)
 443 {
 444     int size, align, r, args_size, i, func_call;
 445     Sym *func_sym;
 446
 447     args_size = 0;
 448     for(i = 0;i < nb_args; i++) {
 449         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 450             size = type_size(&vtop->type, &align);
 451             /* align to stack align size */
 452             size = (size + 3) & ~3;
 453             /* allocate the necessary size on stack */
 454             oad(0xec81, size); /* sub $xxx, %esp */
 455             /* generate structure store */
 456             r = get_reg(RC_INT);
 457             o(0x89); /* mov %esp, r */
 458             o(0xe0 + r);
 459             vset(&vtop->type, r | VT_LVAL, 0);
 460             vswap();
 461             vstore();
 462             args_size += size;
 463         } else if (is_float(vtop->type.t)) {
 464             gv(RC_FLOAT); /* only one float register */
 465             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 466                 size = 4;
 467             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 468                 size = 8;
 469             else
 470                 size = 12;
 471             oad(0xec81, size); /* sub $xxx, %esp */
 472             if (size == 12)
 473                 o(0x7cdb);
 474             else
 475                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 476             g(0x24);
 477             g(0x00);
 478             args_size += size;
 479         } else {
 480             /* simple type (currently always same size) */
 481             /* XXX: implicit cast ? */
 482             r = gv(RC_INT);
 483             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 484                 size = 8;
 485                 o(0x50 + vtop->r2); /* push r */
 486             } else {
 487                 size = 4;
 488             }
 489             o(0x50 + r); /* push r */
 490             args_size += size;
 491         }
 492         vtop--;
 493     }
 494     save_regs(0); /* save used temporary registers */
 495     func_sym = vtop->type.ref;
 496     func_call = func_sym->a.func_call;
 497     /* fast call case */
 498     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 499         func_call == FUNC_FASTCALLW) {
 500         int fastcall_nb_regs;
 501         uint8_t *fastcall_regs_ptr;
 502         if (func_call == FUNC_FASTCALLW) {
 503             fastcall_regs_ptr = fastcallw_regs;
 504             fastcall_nb_regs = 2;
 505         } else {
 506             fastcall_regs_ptr = fastcall_regs;
 507             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 508         }
 509         for(i = 0;i < fastcall_nb_regs; i++) {
 510             if (args_size <= 0)
 511                 break;
 512             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 513             /* XXX: incorrect for struct/floats */
 514             args_size -= 4;
 515         }
 516     }
 517 #ifndef TCC_TARGET_PE
 518     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 519         args_size -= 4;
 520 #endif
 521     gcall_or_jmp(0);
 522
 523     if (args_size && func_call != FUNC_STDCALL)
 524         gadd_sp(args_size);
 525     vtop--;
 526 }
 527
 528 #ifdef TCC_TARGET_PE
 529 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 530 #else
 531 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 532 #endif
 533
 534 /* generate function prolog of type 't' */
 535 ST_FUNC void gfunc_prolog(CType *func_type)
 536 {
 537     int addr, align, size, func_call, fastcall_nb_regs;
 538     int param_index, param_addr;
 539     uint8_t *fastcall_regs_ptr;
 540     Sym *sym;
 541     CType *type;
 542
 543     sym = func_type->ref;
 544     func_call = sym->a.func_call;
 545     addr = 8;
 546     loc = 0;
 547     func_vc = 0;
 548
 549     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 550         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 551         fastcall_regs_ptr = fastcall_regs;
 552     } else if (func_call == FUNC_FASTCALLW) {
 553         fastcall_nb_regs = 2;
 554         fastcall_regs_ptr = fastcallw_regs;
 555     } else {
 556         fastcall_nb_regs = 0;
 557         fastcall_regs_ptr = NULL;
 558     }
 559     param_index = 0;
 560
 561     ind += FUNC_PROLOG_SIZE;
 562     func_sub_sp_offset = ind;
 563     /* if the function returns a structure, then add an
 564        implicit pointer parameter */
 565     func_vt = sym->type;
 566     func_var = (sym->c == FUNC_ELLIPSIS);
 567 #ifdef TCC_TARGET_PE
 568     size = type_size(&func_vt,&align);
 569     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 570 #else
 571     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 572 #endif
 573         /* XXX: fastcall case ? */
 574         func_vc = addr;
 575         addr += 4;
 576         param_index++;
 577     }
 578     /* define parameters */
 579     while ((sym = sym->next) != NULL) {
 580         type = &sym->type;
 581         size = type_size(type, &align);
 582         size = (size + 3) & ~3;
 583 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 584         /* structs are passed as pointer */
 585         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 586             size = 4;
 587         }
 588 #endif
 589         if (param_index < fastcall_nb_regs) {
 590             /* save FASTCALL register */
 591             loc -= 4;
 592             o(0x89);     /* movl */
 593             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 594             param_addr = loc;
 595         } else {
 596             param_addr = addr;
 597             addr += size;
 598         }
 599         sym_push(sym->v & ~SYM_FIELD, type,
 600                  VT_LOCAL | lvalue_type(type->t), param_addr);
 601         param_index++;
 602     }
 603     func_ret_sub = 0;
 604     /* pascal type call ? */
 605     if (func_call == FUNC_STDCALL)
 606         func_ret_sub = addr - 8;
 607 #ifndef TCC_TARGET_PE
 608     else if (func_vc)
 609         func_ret_sub = 4;
 610 #endif
 611
 612 #ifdef CONFIG_TCC_BCHECK
 613     /* leave some room for bound checking code */
 614     if (tcc_state->do_bounds_check) {
 615         func_bound_offset = lbounds_section->data_offset;
 616         func_bound_ind = ind;
 617         oad(0xb8, 0); /* lbound section pointer */
 618         oad(0xb8, 0); /* call to function */
 619     }
 620 #endif
 621 }
 622
 623 /* generate function epilog */
 624 ST_FUNC void gfunc_epilog(void)
 625 {
 626     addr_t v, saved_ind;
 627
 628 #ifdef CONFIG_TCC_BCHECK
 629     if (tcc_state->do_bounds_check
 630      && func_bound_offset != lbounds_section->data_offset) {
 631         addr_t saved_ind;
 632         addr_t *bounds_ptr;
 633         Sym *sym_data;
 634
 635         /* add end of table info */
 636         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 637         *bounds_ptr = 0;
 638
 639         /* generate bound local allocation */
 640         saved_ind = ind;
 641         ind = func_bound_ind;
 642         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 643                                func_bound_offset, lbounds_section->data_offset);
 644         greloc(cur_text_section, sym_data,
 645                ind + 1, R_386_32);
 646         oad(0xb8, 0); /* mov %eax, xxx */
 647         gen_static_call(TOK___bound_local_new);
 648         ind = saved_ind;
 649
 650         /* generate bound check local freeing */
 651         o(0x5250); /* save returned value, if any */
 652         greloc(cur_text_section, sym_data, ind + 1, R_386_32);
 653         oad(0xb8, 0); /* mov %eax, xxx */
 654         gen_static_call(TOK___bound_local_delete);
 655         o(0x585a); /* restore returned value, if any */
 656     }
 657 #endif
 658     o(0x5b * USE_EBX); /* pop ebx */
 659     o(0xc9); /* leave */
 660     if (func_ret_sub == 0) {
 661         o(0xc3); /* ret */
 662     } else {
 663         o(0xc2); /* ret n */
 664         g(func_ret_sub);
 665         g(func_ret_sub >> 8);
 666     }
 667     /* align local size to word & save local variables */
 668
 669     v = (-loc + 3) & -4;
 670     saved_ind = ind;
 671     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 672 #ifdef TCC_TARGET_PE
 673     if (v >= 4096) {
 674         oad(0xb8, v); /* mov stacksize, %eax */
 675         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 676     } else
 677 #endif
 678     {
 679         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 680         o(0xec81);  /* sub esp, stacksize */
 681         gen_le32(v);
 682 #ifdef TCC_TARGET_PE
 683         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 684 #endif
 685     }
 686     o(0x53 * USE_EBX); /* push ebx */
 687     ind = saved_ind;
 688 }
 689
 690 /* generate a jump to a label */
 691 ST_FUNC int gjmp(int t)
 692 {
 693     return psym(0xe9, t);
 694 }
 695
 696 /* generate a jump to a fixed address */
 697 ST_FUNC void gjmp_addr(int a)
 698 {
 699     int r;
 700     r = a - ind - 2;
 701     if (r == (char)r) {
 702         g(0xeb);
 703         g(r);
 704     } else {
 705         oad(0xe9, a - ind - 5);
 706     }
 707 }
 708
 709 ST_FUNC void gtst_addr(int inv, int a)
 710 {
 711     inv ^= (vtop--)->c.i;
 712     a -= ind + 2;
 713     if (a == (char)a) {
 714         g(inv - 32);
 715         g(a);
 716     } else {
 717         g(0x0f);
 718         oad(inv - 16, a - 4);
 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 (v == VT_CMP) {
 727         /* fast case : can jump directly since flags are set */
 728         g(0x0f);
 729         t = psym((vtop->c.i - 16) ^ inv, t);
 730     } else if (v == VT_JMP || v == VT_JMPI) {
 731         /* && or || optimization */
 732         if ((v & 1) == inv) {
 733             /* insert vtop->c jump list in t */
 734             uint32_t n1, n = vtop->c.i;
 735             if (n) {
 736                 while ((n1 = read32le(cur_text_section->data + n)))
 737                     n = n1;
 738                 write32le(cur_text_section->data + n, t);
 739                 t = vtop->c.i;
 740             }
 741         } else {
 742             t = gjmp(t);
 743             gsym(vtop->c.i);
 744         }
 745     }
 746     vtop--;
 747     return t;
 748 }
 749
 750 /* generate an integer binary operation */
 751 ST_FUNC void gen_opi(int op)
 752 {
 753     int r, fr, opc, c;
 754
 755     switch(op) {
 756     case '+':
 757     case TOK_ADDC1: /* add with carry generation */
 758         opc = 0;
 759     gen_op8:
 760         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 761             /* constant case */
 762             vswap();
 763             r = gv(RC_INT);
 764             vswap();
 765             c = vtop->c.i;
 766             if (c == (char)c) {
 767                 /* generate inc and dec for smaller code */
 768                 if (c==1 && opc==0 && op != TOK_ADDC1) {
 769                     o (0x40 | r); // inc
 770                 } else if (c==1 && opc==5 && op != TOK_SUBC1) {
 771                     o (0x48 | r); // dec
 772                 } else {
 773                     o(0x83);
 774                     o(0xc0 | (opc << 3) | r);
 775                     g(c);
 776                 }
 777             } else {
 778                 o(0x81);
 779                 oad(0xc0 | (opc << 3) | r, c);
 780             }
 781         } else {
 782             gv2(RC_INT, RC_INT);
 783             r = vtop[-1].r;
 784             fr = vtop[0].r;
 785             o((opc << 3) | 0x01);
 786             o(0xc0 + r + fr * 8);
 787         }
 788         vtop--;
 789         if (op >= TOK_ULT && op <= TOK_GT) {
 790             vtop->r = VT_CMP;
 791             vtop->c.i = op;
 792         }
 793         break;
 794     case '-':
 795     case TOK_SUBC1: /* sub with carry generation */
 796         opc = 5;
 797         goto gen_op8;
 798     case TOK_ADDC2: /* add with carry use */
 799         opc = 2;
 800         goto gen_op8;
 801     case TOK_SUBC2: /* sub with carry use */
 802         opc = 3;
 803         goto gen_op8;
 804     case '&':
 805         opc = 4;
 806         goto gen_op8;
 807     case '^':
 808         opc = 6;
 809         goto gen_op8;
 810     case '|':
 811         opc = 1;
 812         goto gen_op8;
 813     case '*':
 814         gv2(RC_INT, RC_INT);
 815         r = vtop[-1].r;
 816         fr = vtop[0].r;
 817         vtop--;
 818         o(0xaf0f); /* imul fr, r */
 819         o(0xc0 + fr + r * 8);
 820         break;
 821     case TOK_SHL:
 822         opc = 4;
 823         goto gen_shift;
 824     case TOK_SHR:
 825         opc = 5;
 826         goto gen_shift;
 827     case TOK_SAR:
 828         opc = 7;
 829     gen_shift:
 830         opc = 0xc0 | (opc << 3);
 831         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 832             /* constant case */
 833             vswap();
 834             r = gv(RC_INT);
 835             vswap();
 836             c = vtop->c.i & 0x1f;
 837             o(0xc1); /* shl/shr/sar $xxx, r */
 838             o(opc | r);
 839             g(c);
 840         } else {
 841             /* we generate the shift in ecx */
 842             gv2(RC_INT, RC_ECX);
 843             r = vtop[-1].r;
 844             o(0xd3); /* shl/shr/sar %cl, r */
 845             o(opc | r);
 846         }
 847         vtop--;
 848         break;
 849     case '/':
 850     case TOK_UDIV:
 851     case TOK_PDIV:
 852     case '%':
 853     case TOK_UMOD:
 854     case TOK_UMULL:
 855         /* first operand must be in eax */
 856         /* XXX: need better constraint for second operand */
 857         gv2(RC_EAX, RC_ECX);
 858         r = vtop[-1].r;
 859         fr = vtop[0].r;
 860         vtop--;
 861         save_reg(TREG_EDX);
 862         /* save EAX too if used otherwise */
 863         save_reg_upstack(TREG_EAX, 1);
 864         if (op == TOK_UMULL) {
 865             o(0xf7); /* mul fr */
 866             o(0xe0 + fr);
 867             vtop->r2 = TREG_EDX;
 868             r = TREG_EAX;
 869         } else {
 870             if (op == TOK_UDIV || op == TOK_UMOD) {
 871                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 872                 o(0xf0 + fr);
 873             } else {
 874                 o(0xf799); /* cltd, idiv fr, %eax */
 875                 o(0xf8 + fr);
 876             }
 877             if (op == '%' || op == TOK_UMOD)
 878                 r = TREG_EDX;
 879             else
 880                 r = TREG_EAX;
 881         }
 882         vtop->r = r;
 883         break;
 884     default:
 885         opc = 7;
 886         goto gen_op8;
 887     }
 888 }
 889
 890 /* generate a floating point operation 'v = t1 op t2' instruction. The
 891    two operands are guaranted to have the same floating point type */
 892 /* XXX: need to use ST1 too */
 893 ST_FUNC void gen_opf(int op)
 894 {
 895     int a, ft, fc, swapped, r;
 896
 897     /* convert constants to memory references */
 898     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 899         vswap();
 900         gv(RC_FLOAT);
 901         vswap();
 902     }
 903     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 904         gv(RC_FLOAT);
 905
 906     /* must put at least one value in the floating point register */
 907     if ((vtop[-1].r & VT_LVAL) &&
 908         (vtop[0].r & VT_LVAL)) {
 909         vswap();
 910         gv(RC_FLOAT);
 911         vswap();
 912     }
 913     swapped = 0;
 914     /* swap the stack if needed so that t1 is the register and t2 is
 915        the memory reference */
 916     if (vtop[-1].r & VT_LVAL) {
 917         vswap();
 918         swapped = 1;
 919     }
 920     if (op >= TOK_ULT && op <= TOK_GT) {
 921         /* load on stack second operand */
 922         load(TREG_ST0, vtop);
 923         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 924         if (op == TOK_GE || op == TOK_GT)
 925             swapped = !swapped;
 926         else if (op == TOK_EQ || op == TOK_NE)
 927             swapped = 0;
 928         if (swapped)
 929             o(0xc9d9); /* fxch %st(1) */
 930         if (op == TOK_EQ || op == TOK_NE)
 931             o(0xe9da); /* fucompp */
 932         else
 933             o(0xd9de); /* fcompp */
 934         o(0xe0df); /* fnstsw %ax */
 935         if (op == TOK_EQ) {
 936             o(0x45e480); /* and $0x45, %ah */
 937             o(0x40fC80); /* cmp $0x40, %ah */
 938         } else if (op == TOK_NE) {
 939             o(0x45e480); /* and $0x45, %ah */
 940             o(0x40f480); /* xor $0x40, %ah */
 941             op = TOK_NE;
 942         } else if (op == TOK_GE || op == TOK_LE) {
 943             o(0x05c4f6); /* test $0x05, %ah */
 944             op = TOK_EQ;
 945         } else {
 946             o(0x45c4f6); /* test $0x45, %ah */
 947             op = TOK_EQ;
 948         }
 949         vtop--;
 950         vtop->r = VT_CMP;
 951         vtop->c.i = op;
 952     } else {
 953         /* no memory reference possible for long double operations */
 954         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 955             load(TREG_ST0, vtop);
 956             swapped = !swapped;
 957         }
 958
 959         switch(op) {
 960         default:
 961         case '+':
 962             a = 0;
 963             break;
 964         case '-':
 965             a = 4;
 966             if (swapped)
 967                 a++;
 968             break;
 969         case '*':
 970             a = 1;
 971             break;
 972         case '/':
 973             a = 6;
 974             if (swapped)
 975                 a++;
 976             break;
 977         }
 978         ft = vtop->type.t;
 979         fc = vtop->c.i;
 980         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 981             o(0xde); /* fxxxp %st, %st(1) */
 982             o(0xc1 + (a << 3));
 983         } else {
 984             /* if saved lvalue, then we must reload it */
 985             r = vtop->r;
 986             if ((r & VT_VALMASK) == VT_LLOCAL) {
 987                 SValue v1;
 988                 r = get_reg(RC_INT);
 989                 v1.type.t = VT_INT;
 990                 v1.r = VT_LOCAL | VT_LVAL;
 991                 v1.c.i = fc;
 992                 load(r, &v1);
 993                 fc = 0;
 994             }
 995
 996             if ((ft & VT_BTYPE) == VT_DOUBLE)
 997                 o(0xdc);
 998             else
 999                 o(0xd8);
1000             gen_modrm(a, r, vtop->sym, fc);
1001         }
1002         vtop--;
1003     }
1004 }
1005
1006 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1007    and 'long long' cases. */
1008 ST_FUNC void gen_cvt_itof(int t)
1009 {
1010     save_reg(TREG_ST0);
1011     gv(RC_INT);
1012     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
1013         /* signed long long to float/double/long double (unsigned case
1014            is handled generically) */
1015         o(0x50 + vtop->r2); /* push r2 */
1016         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1017         o(0x242cdf); /* fildll (%esp) */
1018         o(0x08c483); /* add $8, %esp */
1019     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
1020                (VT_INT | VT_UNSIGNED)) {
1021         /* unsigned int to float/double/long double */
1022         o(0x6a); /* push $0 */
1023         g(0x00);
1024         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1025         o(0x242cdf); /* fildll (%esp) */
1026         o(0x08c483); /* add $8, %esp */
1027     } else {
1028         /* int to float/double/long double */
1029         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1030         o(0x2404db); /* fildl (%esp) */
1031         o(0x04c483); /* add $4, %esp */
1032     }
1033     vtop->r = TREG_ST0;
1034 }
1035
1036 /* convert fp to int 't' type */
1037 ST_FUNC void gen_cvt_ftoi(int t)
1038 {
1039 #if 1
1040     gv(RC_FLOAT);
1041     save_reg(TREG_EAX);
1042     save_reg(TREG_EDX);
1043     gen_static_call(TOK___tcc_cvt_ftol);
1044     vtop->r = TREG_EAX; /* mark reg as used */
1045     if (t == VT_LLONG)
1046         vtop->r2 = TREG_EDX;
1047 #else
1048     int bt = vtop->type.t & VT_BTYPE;
1049     if (bt == VT_FLOAT)
1050         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1051     else if (bt == VT_LDOUBLE)
1052         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1053     else
1054         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1055     vswap();
1056     gfunc_call(1);
1057     vpushi(0);
1058     vtop->r = REG_IRET;
1059     vtop->r2 = REG_LRET;
1060 #endif
1061 }
1062
1063 /* convert from one floating point type to another */
1064 ST_FUNC void gen_cvt_ftof(int t)
1065 {
1066     /* all we have to do on i386 is to put the float in a register */
1067     gv(RC_FLOAT);
1068 }
1069
1070 /* computed goto support */
1071 ST_FUNC void ggoto(void)
1072 {
1073     gcall_or_jmp(1);
1074     vtop--;
1075 }
1076
1077 /* bound check support functions */
1078 #ifdef CONFIG_TCC_BCHECK
1079
1080 /* generate a bounded pointer addition */
1081 ST_FUNC void gen_bounded_ptr_add(void)
1082 {
1083     /* prepare fast i386 function call (args in eax and edx) */
1084     gv2(RC_EAX, RC_EDX);
1085     /* save all temporary registers */
1086     vtop -= 2;
1087     save_regs(0);
1088     /* do a fast function call */
1089     gen_static_call(TOK___bound_ptr_add);
1090     /* returned pointer is in eax */
1091     vtop++;
1092     vtop->r = TREG_EAX | VT_BOUNDED;
1093     /* address of bounding function call point */
1094     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1095 }
1096
1097 /* patch pointer addition in vtop so that pointer dereferencing is
1098    also tested */
1099 ST_FUNC void gen_bounded_ptr_deref(void)
1100 {
1101     addr_t func;
1102     int  size, align;
1103     Elf32_Rel *rel;
1104     Sym *sym;
1105
1106     size = 0;
1107     /* XXX: put that code in generic part of tcc */
1108     if (!is_float(vtop->type.t)) {
1109         if (vtop->r & VT_LVAL_BYTE)
1110             size = 1;
1111         else if (vtop->r & VT_LVAL_SHORT)
1112             size = 2;
1113     }
1114     if (!size)
1115         size = type_size(&vtop->type, &align);
1116     switch(size) {
1117     case  1: func = TOK___bound_ptr_indir1; break;
1118     case  2: func = TOK___bound_ptr_indir2; break;
1119     case  4: func = TOK___bound_ptr_indir4; break;
1120     case  8: func = TOK___bound_ptr_indir8; break;
1121     case 12: func = TOK___bound_ptr_indir12; break;
1122     case 16: func = TOK___bound_ptr_indir16; break;
1123     default:
1124         tcc_error("unhandled size when dereferencing bounded pointer");
1125         func = 0;
1126         break;
1127     }
1128
1129     /* patch relocation */
1130     /* XXX: find a better solution ? */
1131     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1132     sym = external_global_sym(func, &func_old_type, 0);
1133     if (!sym->c)
1134         put_extern_sym(sym, NULL, 0, 0);
1135     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1136 }
1137 #endif
1138
1139 /* Save the stack pointer onto the stack */
1140 ST_FUNC void gen_vla_sp_save(int addr) {
1141     /* mov %esp,addr(%ebp)*/
1142     o(0x89);
1143     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1144 }
1145
1146 /* Restore the SP from a location on the stack */
1147 ST_FUNC void gen_vla_sp_restore(int addr) {
1148     o(0x8b);
1149     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1150 }
1151
1152 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1153 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1154 #ifdef TCC_TARGET_PE
1155     /* alloca does more than just adjust %rsp on Windows */
1156     vpush_global_sym(&func_old_type, TOK_alloca);
1157     vswap(); /* Move alloca ref past allocation size */
1158     gfunc_call(1);
1159 #else
1160     int r;
1161     r = gv(RC_INT); /* allocation size */
1162     /* sub r,%rsp */
1163     o(0x2b);
1164     o(0xe0 | r);
1165     /* We align to 16 bytes rather than align */
1166     /* and ~15, %esp */
1167     o(0xf0e483);
1168     vpop();
1169 #endif
1170 }
1171
1172 /* end of X86 code generator */
1173 /*************************************************************/
1174 #endif
1175 /*************************************************************/