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         4
  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_IRET    RC_EAX /* function return: integer register */
  37 #define RC_LRET    RC_EDX /* function return: second integer register */
  38 #define RC_FRET    RC_ST0 /* function return: float register */
  39
  40 /* pretty names for the registers */
  41 enum {
  42     TREG_EAX = 0,
  43     TREG_ECX,
  44     TREG_EDX,
  45     TREG_ST0,
  46     TREG_ESP = 4
  47 };
  48
  49 /* return registers for function */
  50 #define REG_IRET TREG_EAX /* single word int return register */
  51 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  52 #define REG_FRET TREG_ST0 /* float return register */
  53
  54 /* defined if function parameters must be evaluated in reverse order */
  55 #define INVERT_FUNC_PARAMS
  56
  57 /* defined if structures are passed as pointers. Otherwise structures
  58    are directly pushed on stack. */
  59 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  60
  61 /* pointer size, in bytes */
  62 #define PTR_SIZE 4
  63
  64 /* long double size and alignment, in bytes */
  65 #define LDOUBLE_SIZE  12
  66 #define LDOUBLE_ALIGN 4
  67 /* maximum alignment (for aligned attribute support) */
  68 #define MAX_ALIGN     8
  69
  70
  71 #define psym oad
  72
  73 /******************************************************/
  74 /* ELF defines */
  75
  76 #define EM_TCC_TARGET EM_386
  77
  78 /* relocation type for 32 bit data relocation */
  79 #define R_DATA_32   R_386_32
  80 #define R_DATA_PTR  R_386_32
  81 #define R_JMP_SLOT  R_386_JMP_SLOT
  82 #define R_COPY      R_386_COPY
  83
  84 #define ELF_START_ADDR 0x08048000
  85 #define ELF_PAGE_SIZE  0x1000
  86
  87 /******************************************************/
  88 #else /* ! TARGET_DEFS_ONLY */
  89 /******************************************************/
  90 #include "tcc.h"
  91
  92 ST_DATA const int reg_classes[NB_REGS] = {
  93     /* eax */ RC_INT | RC_EAX,
  94     /* ecx */ RC_INT | RC_ECX,
  95     /* edx */ RC_INT | RC_EDX,
  96     /* st0 */ RC_FLOAT | RC_ST0,
  97 };
  98
  99 static unsigned long func_sub_sp_offset;
 100 static int func_ret_sub;
 101 #ifdef CONFIG_TCC_BCHECK
 102 static addr_t func_bound_offset;
 103 #endif
 104
 105 /* XXX: make it faster ? */
 106 ST_FUNC void g(int c)
 107 {
 108     int ind1;
 109     ind1 = ind + 1;
 110     if (ind1 > cur_text_section->data_allocated)
 111         section_realloc(cur_text_section, ind1);
 112     cur_text_section->data[ind] = c;
 113     ind = ind1;
 114 }
 115
 116 ST_FUNC void o(unsigned int c)
 117 {
 118     while (c) {
 119         g(c);
 120         c = c >> 8;
 121     }
 122 }
 123
 124 ST_FUNC void gen_le16(int v)
 125 {
 126     g(v);
 127     g(v >> 8);
 128 }
 129
 130 ST_FUNC void gen_le32(int c)
 131 {
 132     g(c);
 133     g(c >> 8);
 134     g(c >> 16);
 135     g(c >> 24);
 136 }
 137
 138 /* output a symbol and patch all calls to it */
 139 ST_FUNC void gsym_addr(int t, int a)
 140 {
 141     while (t) {
 142         unsigned char *ptr = cur_text_section->data + t;
 143         uint32_t n = read32le(ptr); /* next value */
 144         write32le(ptr, a - t - 4);
 145         t = n;
 146     }
 147 }
 148
 149 ST_FUNC void gsym(int t)
 150 {
 151     gsym_addr(t, ind);
 152 }
 153
 154 /* psym is used to put an instruction with a data field which is a
 155    reference to a symbol. It is in fact the same as oad ! */
 156 #define psym oad
 157
 158 /* instruction + 4 bytes data. Return the address of the data */
 159 ST_FUNC int oad(int c, int s)
 160 {
 161     int ind1;
 162
 163     o(c);
 164     ind1 = ind + 4;
 165     if (ind1 > cur_text_section->data_allocated)
 166         section_realloc(cur_text_section, ind1);
 167     write32le(cur_text_section->data + ind, s);
 168     s = ind;
 169     ind = ind1;
 170     return s;
 171 }
 172
 173 /* output constant with relocation if 'r & VT_SYM' is true */
 174 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 175 {
 176     if (r & VT_SYM)
 177         greloc(cur_text_section, sym, ind, R_386_32);
 178     gen_le32(c);
 179 }
 180
 181 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 182 {
 183     if (r & VT_SYM)
 184         greloc(cur_text_section, sym, ind, R_386_PC32);
 185     gen_le32(c - 4);
 186 }
 187
 188 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 189    opcode bits */
 190 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 191 {
 192     op_reg = op_reg << 3;
 193     if ((r & VT_VALMASK) == VT_CONST) {
 194         /* constant memory reference */
 195         o(0x05 | op_reg);
 196         gen_addr32(r, sym, c);
 197     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 198         /* currently, we use only ebp as base */
 199         if (c == (char)c) {
 200             /* short reference */
 201             o(0x45 | op_reg);
 202             g(c);
 203         } else {
 204             oad(0x85 | op_reg, c);
 205         }
 206     } else {
 207         g(0x00 | op_reg | (r & VT_VALMASK));
 208     }
 209 }
 210
 211 /* load 'r' from value 'sv' */
 212 ST_FUNC void load(int r, SValue *sv)
 213 {
 214     int v, t, ft, fc, fr;
 215     SValue v1;
 216
 217 #ifdef TCC_TARGET_PE
 218     SValue v2;
 219     sv = pe_getimport(sv, &v2);
 220 #endif
 221
 222     fr = sv->r;
 223     ft = sv->type.t;
 224     fc = sv->c.i;
 225
 226     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 227
 228     v = fr & VT_VALMASK;
 229     if (fr & VT_LVAL) {
 230         if (v == VT_LLOCAL) {
 231             v1.type.t = VT_INT;
 232             v1.r = VT_LOCAL | VT_LVAL;
 233             v1.c.i = fc;
 234             fr = r;
 235             if (!(reg_classes[fr] & RC_INT))
 236                 fr = get_reg(RC_INT);
 237             load(fr, &v1);
 238         }
 239         if ((ft & VT_BTYPE) == VT_FLOAT) {
 240             o(0xd9); /* flds */
 241             r = 0;
 242         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 243             o(0xdd); /* fldl */
 244             r = 0;
 245         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 246             o(0xdb); /* fldt */
 247             r = 5;
 248         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 249             o(0xbe0f);   /* movsbl */
 250         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 251             o(0xb60f);   /* movzbl */
 252         } else if ((ft & VT_TYPE) == VT_SHORT) {
 253             o(0xbf0f);   /* movswl */
 254         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 255             o(0xb70f);   /* movzwl */
 256         } else {
 257             o(0x8b);     /* movl */
 258         }
 259         gen_modrm(r, fr, sv->sym, fc);
 260     } else {
 261         if (v == VT_CONST) {
 262             o(0xb8 + r); /* mov $xx, r */
 263             gen_addr32(fr, sv->sym, fc);
 264         } else if (v == VT_LOCAL) {
 265             if (fc) {
 266                 o(0x8d); /* lea xxx(%ebp), r */
 267                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 268             } else {
 269                 o(0x89);
 270                 o(0xe8 + r); /* mov %ebp, r */
 271             }
 272         } else if (v == VT_CMP) {
 273             oad(0xb8 + r, 0); /* mov $0, r */
 274             o(0x0f); /* setxx %br */
 275             o(fc);
 276             o(0xc0 + r);
 277         } else if (v == VT_JMP || v == VT_JMPI) {
 278             t = v & 1;
 279             oad(0xb8 + r, t); /* mov $1, r */
 280             o(0x05eb); /* jmp after */
 281             gsym(fc);
 282             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 283         } else if (v != r) {
 284             o(0x89);
 285             o(0xc0 + r + v * 8); /* mov v, r */
 286         }
 287     }
 288 }
 289
 290 /* store register 'r' in lvalue 'v' */
 291 ST_FUNC void store(int r, SValue *v)
 292 {
 293     int fr, bt, ft, fc;
 294
 295 #ifdef TCC_TARGET_PE
 296     SValue v2;
 297     v = pe_getimport(v, &v2);
 298 #endif
 299
 300     ft = v->type.t;
 301     fc = v->c.i;
 302     fr = v->r & VT_VALMASK;
 303     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 304     bt = ft & VT_BTYPE;
 305     /* XXX: incorrect if float reg to reg */
 306     if (bt == VT_FLOAT) {
 307         o(0xd9); /* fsts */
 308         r = 2;
 309     } else if (bt == VT_DOUBLE) {
 310         o(0xdd); /* fstpl */
 311         r = 2;
 312     } else if (bt == VT_LDOUBLE) {
 313         o(0xc0d9); /* fld %st(0) */
 314         o(0xdb); /* fstpt */
 315         r = 7;
 316     } else {
 317         if (bt == VT_SHORT)
 318             o(0x66);
 319         if (bt == VT_BYTE || bt == VT_BOOL)
 320             o(0x88);
 321         else
 322             o(0x89);
 323     }
 324     if (fr == VT_CONST ||
 325         fr == VT_LOCAL ||
 326         (v->r & VT_LVAL)) {
 327         gen_modrm(r, v->r, v->sym, fc);
 328     } else if (fr != r) {
 329         o(0xc0 + fr + r * 8); /* mov r, fr */
 330     }
 331 }
 332
 333 static void gadd_sp(int val)
 334 {
 335     if (val == (char)val) {
 336         o(0xc483);
 337         g(val);
 338     } else {
 339         oad(0xc481, val); /* add $xxx, %esp */
 340     }
 341 }
 342
 343 static void gen_static_call(int v)
 344 {
 345     Sym *sym;
 346
 347     sym = external_global_sym(v, &func_old_type, 0);
 348     oad(0xe8, -4);
 349     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 350 }
 351
 352 /* 'is_jmp' is '1' if it is a jump */
 353 static void gcall_or_jmp(int is_jmp)
 354 {
 355     int r;
 356     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 357         int rt;
 358         /* constant case */
 359         if (vtop->r & VT_SYM) {
 360             /* relocation case */
 361             greloc(cur_text_section, vtop->sym,
 362                    ind + 1, R_386_PC32);
 363         } else {
 364             /* put an empty PC32 relocation */
 365             put_elf_reloc(symtab_section, cur_text_section,
 366                           ind + 1, R_386_PC32, 0);
 367         }
 368         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 369         /* extend the return value to the whole register if necessary
 370            visual studio and gcc do not always set the whole eax register
 371            when assigning the return value of a function  */
 372         rt = vtop->type.ref->type.t;
 373         switch (rt & VT_BTYPE) {
 374             case VT_BYTE:
 375                 if (rt & VT_UNSIGNED) {
 376                     o(0xc0b60f); /* movzx %al, %eax */
 377                 }
 378                 else {
 379                     o(0xc0be0f); /* movsx %al, %eax */
 380                 }
 381                 break;
 382             case VT_SHORT:
 383                 if (rt & VT_UNSIGNED) {
 384                     o(0xc0b70f); /* movzx %ax, %eax */
 385                 }
 386                 else {
 387                     o(0xc0bf0f); /* movsx %ax, %eax */
 388                 }
 389                 break;
 390             default:
 391                 break;
 392         }
 393     } else {
 394         /* otherwise, indirect call */
 395         r = gv(RC_INT);
 396         o(0xff); /* call/jmp *r */
 397         o(0xd0 + r + (is_jmp << 4));
 398     }
 399 }
 400
 401 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 402 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 403
 404 /* Return the number of registers needed to return the struct, or 0 if
 405    returning via struct pointer. */
 406 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 407 {
 408 #ifdef TCC_TARGET_PE
 409     int size, align;
 410
 411     *ret_align = 1; // Never have to re-align return values for x86
 412     *regsize = 4;
 413     size = type_size(vt, &align);
 414     if (size > 8) {
 415         return 0;
 416     } else if (size > 4) {
 417         ret->ref = NULL;
 418         ret->t = VT_LLONG;
 419         return 1;
 420     } else {
 421         ret->ref = NULL;
 422         ret->t = VT_INT;
 423         return 1;
 424     }
 425 #else
 426     *ret_align = 1; // Never have to re-align return values for x86
 427     return 0;
 428 #endif
 429 }
 430
 431 /* Generate function call. The function address is pushed first, then
 432    all the parameters in call order. This functions pops all the
 433    parameters and the function address. */
 434 ST_FUNC void gfunc_call(int nb_args)
 435 {
 436     int size, align, r, args_size, i, func_call;
 437     Sym *func_sym;
 438
 439     args_size = 0;
 440     for(i = 0;i < nb_args; i++) {
 441         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 442             size = type_size(&vtop->type, &align);
 443             /* align to stack align size */
 444             size = (size + 3) & ~3;
 445             /* allocate the necessary size on stack */
 446             oad(0xec81, size); /* sub $xxx, %esp */
 447             /* generate structure store */
 448             r = get_reg(RC_INT);
 449             o(0x89); /* mov %esp, r */
 450             o(0xe0 + r);
 451             vset(&vtop->type, r | VT_LVAL, 0);
 452             vswap();
 453             vstore();
 454             args_size += size;
 455         } else if (is_float(vtop->type.t)) {
 456             gv(RC_FLOAT); /* only one float register */
 457             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 458                 size = 4;
 459             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 460                 size = 8;
 461             else
 462                 size = 12;
 463             oad(0xec81, size); /* sub $xxx, %esp */
 464             if (size == 12)
 465                 o(0x7cdb);
 466             else
 467                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 468             g(0x24);
 469             g(0x00);
 470             args_size += size;
 471         } else {
 472             /* simple type (currently always same size) */
 473             /* XXX: implicit cast ? */
 474             r = gv(RC_INT);
 475             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 476                 size = 8;
 477                 o(0x50 + vtop->r2); /* push r */
 478             } else {
 479                 size = 4;
 480             }
 481             o(0x50 + r); /* push r */
 482             args_size += size;
 483         }
 484         vtop--;
 485     }
 486     save_regs(0); /* save used temporary registers */
 487     func_sym = vtop->type.ref;
 488     func_call = func_sym->a.func_call;
 489     /* fast call case */
 490     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 491         func_call == FUNC_FASTCALLW) {
 492         int fastcall_nb_regs;
 493         uint8_t *fastcall_regs_ptr;
 494         if (func_call == FUNC_FASTCALLW) {
 495             fastcall_regs_ptr = fastcallw_regs;
 496             fastcall_nb_regs = 2;
 497         } else {
 498             fastcall_regs_ptr = fastcall_regs;
 499             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 500         }
 501         for(i = 0;i < fastcall_nb_regs; i++) {
 502             if (args_size <= 0)
 503                 break;
 504             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 505             /* XXX: incorrect for struct/floats */
 506             args_size -= 4;
 507         }
 508     }
 509 #ifndef TCC_TARGET_PE
 510     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 511         args_size -= 4;
 512 #endif
 513     gcall_or_jmp(0);
 514
 515     if (args_size && func_call != FUNC_STDCALL)
 516         gadd_sp(args_size);
 517     vtop--;
 518 }
 519
 520 #ifdef TCC_TARGET_PE
 521 #define FUNC_PROLOG_SIZE 10
 522 #else
 523 #define FUNC_PROLOG_SIZE 9
 524 #endif
 525
 526 /* generate function prolog of type 't' */
 527 ST_FUNC void gfunc_prolog(CType *func_type)
 528 {
 529     int addr, align, size, func_call, fastcall_nb_regs;
 530     int param_index, param_addr;
 531     uint8_t *fastcall_regs_ptr;
 532     Sym *sym;
 533     CType *type;
 534
 535     sym = func_type->ref;
 536     func_call = sym->a.func_call;
 537     addr = 8;
 538     loc = 0;
 539     func_vc = 0;
 540
 541     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 542         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 543         fastcall_regs_ptr = fastcall_regs;
 544     } else if (func_call == FUNC_FASTCALLW) {
 545         fastcall_nb_regs = 2;
 546         fastcall_regs_ptr = fastcallw_regs;
 547     } else {
 548         fastcall_nb_regs = 0;
 549         fastcall_regs_ptr = NULL;
 550     }
 551     param_index = 0;
 552
 553     ind += FUNC_PROLOG_SIZE;
 554     func_sub_sp_offset = ind;
 555     /* if the function returns a structure, then add an
 556        implicit pointer parameter */
 557     func_vt = sym->type;
 558     func_var = (sym->c == FUNC_ELLIPSIS);
 559 #ifdef TCC_TARGET_PE
 560     size = type_size(&func_vt,&align);
 561     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 562 #else
 563     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 564 #endif
 565         /* XXX: fastcall case ? */
 566         func_vc = addr;
 567         addr += 4;
 568         param_index++;
 569     }
 570     /* define parameters */
 571     while ((sym = sym->next) != NULL) {
 572         type = &sym->type;
 573         size = type_size(type, &align);
 574         size = (size + 3) & ~3;
 575 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 576         /* structs are passed as pointer */
 577         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 578             size = 4;
 579         }
 580 #endif
 581         if (param_index < fastcall_nb_regs) {
 582             /* save FASTCALL register */
 583             loc -= 4;
 584             o(0x89);     /* movl */
 585             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 586             param_addr = loc;
 587         } else {
 588             param_addr = addr;
 589             addr += size;
 590         }
 591         sym_push(sym->v & ~SYM_FIELD, type,
 592                  VT_LOCAL | lvalue_type(type->t), param_addr);
 593         param_index++;
 594     }
 595     func_ret_sub = 0;
 596     /* pascal type call ? */
 597     if (func_call == FUNC_STDCALL)
 598         func_ret_sub = addr - 8;
 599 #ifndef TCC_TARGET_PE
 600     else if (func_vc)
 601         func_ret_sub = 4;
 602 #endif
 603
 604 #ifdef CONFIG_TCC_BCHECK
 605     /* leave some room for bound checking code */
 606     if (tcc_state->do_bounds_check) {
 607         oad(0xb8, 0); /* lbound section pointer */
 608         oad(0xb8, 0); /* call to function */
 609         func_bound_offset = lbounds_section->data_offset;
 610     }
 611 #endif
 612 }
 613
 614 /* generate function epilog */
 615 ST_FUNC void gfunc_epilog(void)
 616 {
 617     addr_t v, saved_ind;
 618
 619 #ifdef CONFIG_TCC_BCHECK
 620     if (tcc_state->do_bounds_check
 621      && func_bound_offset != lbounds_section->data_offset) {
 622         addr_t saved_ind;
 623         addr_t *bounds_ptr;
 624         Sym *sym_data;
 625         /* add end of table info */
 626         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 627         *bounds_ptr = 0;
 628         /* generate bound local allocation */
 629         saved_ind = ind;
 630         ind = func_sub_sp_offset;
 631         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 632                                func_bound_offset, lbounds_section->data_offset);
 633         greloc(cur_text_section, sym_data,
 634                ind + 1, R_386_32);
 635         oad(0xb8, 0); /* mov %eax, xxx */
 636         gen_static_call(TOK___bound_local_new);
 637
 638         ind = saved_ind;
 639         /* generate bound check local freeing */
 640         o(0x5250); /* save returned value, if any */
 641         greloc(cur_text_section, sym_data,
 642                ind + 1, R_386_32);
 643         oad(0xb8, 0); /* mov %eax, xxx */
 644         gen_static_call(TOK___bound_local_delete);
 645
 646         o(0x585a); /* restore returned value, if any */
 647     }
 648 #endif
 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     /* align local size to word & save local variables */
 658
 659     v = (-loc + 3) & -4;
 660     saved_ind = ind;
 661     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 662 #ifdef TCC_TARGET_PE
 663     if (v >= 4096) {
 664         oad(0xb8, v); /* mov stacksize, %eax */
 665         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 666     } else
 667 #endif
 668     {
 669         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 670         o(0xec81);  /* sub esp, stacksize */
 671         gen_le32(v);
 672 #if FUNC_PROLOG_SIZE == 10
 673         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 674 #endif
 675     }
 676     ind = saved_ind;
 677 }
 678
 679 /* generate a jump to a label */
 680 ST_FUNC int gjmp(int t)
 681 {
 682     return psym(0xe9, t);
 683 }
 684
 685 /* generate a jump to a fixed address */
 686 ST_FUNC void gjmp_addr(int a)
 687 {
 688     int r;
 689     r = a - ind - 2;
 690     if (r == (char)r) {
 691         g(0xeb);
 692         g(r);
 693     } else {
 694         oad(0xe9, a - ind - 5);
 695     }
 696 }
 697
 698 ST_FUNC void gtst_addr(int inv, int a)
 699 {
 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 }
 710
 711 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 712 ST_FUNC int gtst(int inv, int t)
 713 {
 714     int v = vtop->r & VT_VALMASK;
 715     if (v == VT_CMP) {
 716         /* fast case : can jump directly since flags are set */
 717         g(0x0f);
 718         t = psym((vtop->c.i - 16) ^ inv, t);
 719     } else if (v == VT_JMP || v == VT_JMPI) {
 720         /* && or || optimization */
 721         if ((v & 1) == inv) {
 722             /* insert vtop->c jump list in t */
 723             uint32_t n1, n = vtop->c.i;
 724             if (n) {
 725                 while ((n1 = read32le(cur_text_section->data + n)))
 726                     n = n1;
 727                 write32le(cur_text_section->data + n, t);
 728                 t = vtop->c.i;
 729             }
 730         } else {
 731             t = gjmp(t);
 732             gsym(vtop->c.i);
 733         }
 734     }
 735     vtop--;
 736     return t;
 737 }
 738
 739 /* generate an integer binary operation */
 740 ST_FUNC void gen_opi(int op)
 741 {
 742     int r, fr, opc, c;
 743
 744     switch(op) {
 745     case '+':
 746     case TOK_ADDC1: /* add with carry generation */
 747         opc = 0;
 748     gen_op8:
 749         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 750             /* constant case */
 751             vswap();
 752             r = gv(RC_INT);
 753             vswap();
 754             c = vtop->c.i;
 755             if (c == (char)c) {
 756                 /* generate inc and dec for smaller code */
 757                 if (c==1 && opc==0) {
 758                     o (0x40 | r); // inc
 759                 } else if (c==1 && opc==5) {
 760                     o (0x48 | r); // dec
 761                 } else {
 762                     o(0x83);
 763                     o(0xc0 | (opc << 3) | r);
 764                     g(c);
 765                 }
 766             } else {
 767                 o(0x81);
 768                 oad(0xc0 | (opc << 3) | r, c);
 769             }
 770         } else {
 771             gv2(RC_INT, RC_INT);
 772             r = vtop[-1].r;
 773             fr = vtop[0].r;
 774             o((opc << 3) | 0x01);
 775             o(0xc0 + r + fr * 8);
 776         }
 777         vtop--;
 778         if (op >= TOK_ULT && op <= TOK_GT) {
 779             vtop->r = VT_CMP;
 780             vtop->c.i = op;
 781         }
 782         break;
 783     case '-':
 784     case TOK_SUBC1: /* sub with carry generation */
 785         opc = 5;
 786         goto gen_op8;
 787     case TOK_ADDC2: /* add with carry use */
 788         opc = 2;
 789         goto gen_op8;
 790     case TOK_SUBC2: /* sub with carry use */
 791         opc = 3;
 792         goto gen_op8;
 793     case '&':
 794         opc = 4;
 795         goto gen_op8;
 796     case '^':
 797         opc = 6;
 798         goto gen_op8;
 799     case '|':
 800         opc = 1;
 801         goto gen_op8;
 802     case '*':
 803         gv2(RC_INT, RC_INT);
 804         r = vtop[-1].r;
 805         fr = vtop[0].r;
 806         vtop--;
 807         o(0xaf0f); /* imul fr, r */
 808         o(0xc0 + fr + r * 8);
 809         break;
 810     case TOK_SHL:
 811         opc = 4;
 812         goto gen_shift;
 813     case TOK_SHR:
 814         opc = 5;
 815         goto gen_shift;
 816     case TOK_SAR:
 817         opc = 7;
 818     gen_shift:
 819         opc = 0xc0 | (opc << 3);
 820         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 821             /* constant case */
 822             vswap();
 823             r = gv(RC_INT);
 824             vswap();
 825             c = vtop->c.i & 0x1f;
 826             o(0xc1); /* shl/shr/sar $xxx, r */
 827             o(opc | r);
 828             g(c);
 829         } else {
 830             /* we generate the shift in ecx */
 831             gv2(RC_INT, RC_ECX);
 832             r = vtop[-1].r;
 833             o(0xd3); /* shl/shr/sar %cl, r */
 834             o(opc | r);
 835         }
 836         vtop--;
 837         break;
 838     case '/':
 839     case TOK_UDIV:
 840     case TOK_PDIV:
 841     case '%':
 842     case TOK_UMOD:
 843     case TOK_UMULL:
 844         /* first operand must be in eax */
 845         /* XXX: need better constraint for second operand */
 846         gv2(RC_EAX, RC_ECX);
 847         r = vtop[-1].r;
 848         fr = vtop[0].r;
 849         vtop--;
 850         save_reg(TREG_EDX);
 851         /* save EAX too if used otherwise */
 852         save_reg_upstack(TREG_EAX, 1);
 853         if (op == TOK_UMULL) {
 854             o(0xf7); /* mul fr */
 855             o(0xe0 + fr);
 856             vtop->r2 = TREG_EDX;
 857             r = TREG_EAX;
 858         } else {
 859             if (op == TOK_UDIV || op == TOK_UMOD) {
 860                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 861                 o(0xf0 + fr);
 862             } else {
 863                 o(0xf799); /* cltd, idiv fr, %eax */
 864                 o(0xf8 + fr);
 865             }
 866             if (op == '%' || op == TOK_UMOD)
 867                 r = TREG_EDX;
 868             else
 869                 r = TREG_EAX;
 870         }
 871         vtop->r = r;
 872         break;
 873     default:
 874         opc = 7;
 875         goto gen_op8;
 876     }
 877 }
 878
 879 /* generate a floating point operation 'v = t1 op t2' instruction. The
 880    two operands are guaranted to have the same floating point type */
 881 /* XXX: need to use ST1 too */
 882 ST_FUNC void gen_opf(int op)
 883 {
 884     int a, ft, fc, swapped, r;
 885
 886     /* convert constants to memory references */
 887     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 888         vswap();
 889         gv(RC_FLOAT);
 890         vswap();
 891     }
 892     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 893         gv(RC_FLOAT);
 894
 895     /* must put at least one value in the floating point register */
 896     if ((vtop[-1].r & VT_LVAL) &&
 897         (vtop[0].r & VT_LVAL)) {
 898         vswap();
 899         gv(RC_FLOAT);
 900         vswap();
 901     }
 902     swapped = 0;
 903     /* swap the stack if needed so that t1 is the register and t2 is
 904        the memory reference */
 905     if (vtop[-1].r & VT_LVAL) {
 906         vswap();
 907         swapped = 1;
 908     }
 909     if (op >= TOK_ULT && op <= TOK_GT) {
 910         /* load on stack second operand */
 911         load(TREG_ST0, vtop);
 912         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 913         if (op == TOK_GE || op == TOK_GT)
 914             swapped = !swapped;
 915         else if (op == TOK_EQ || op == TOK_NE)
 916             swapped = 0;
 917         if (swapped)
 918             o(0xc9d9); /* fxch %st(1) */
 919         if (op == TOK_EQ || op == TOK_NE)
 920             o(0xe9da); /* fucompp */
 921         else
 922             o(0xd9de); /* fcompp */
 923         o(0xe0df); /* fnstsw %ax */
 924         if (op == TOK_EQ) {
 925             o(0x45e480); /* and $0x45, %ah */
 926             o(0x40fC80); /* cmp $0x40, %ah */
 927         } else if (op == TOK_NE) {
 928             o(0x45e480); /* and $0x45, %ah */
 929             o(0x40f480); /* xor $0x40, %ah */
 930             op = TOK_NE;
 931         } else if (op == TOK_GE || op == TOK_LE) {
 932             o(0x05c4f6); /* test $0x05, %ah */
 933             op = TOK_EQ;
 934         } else {
 935             o(0x45c4f6); /* test $0x45, %ah */
 936             op = TOK_EQ;
 937         }
 938         vtop--;
 939         vtop->r = VT_CMP;
 940         vtop->c.i = op;
 941     } else {
 942         /* no memory reference possible for long double operations */
 943         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 944             load(TREG_ST0, vtop);
 945             swapped = !swapped;
 946         }
 947
 948         switch(op) {
 949         default:
 950         case '+':
 951             a = 0;
 952             break;
 953         case '-':
 954             a = 4;
 955             if (swapped)
 956                 a++;
 957             break;
 958         case '*':
 959             a = 1;
 960             break;
 961         case '/':
 962             a = 6;
 963             if (swapped)
 964                 a++;
 965             break;
 966         }
 967         ft = vtop->type.t;
 968         fc = vtop->c.i;
 969         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 970             o(0xde); /* fxxxp %st, %st(1) */
 971             o(0xc1 + (a << 3));
 972         } else {
 973             /* if saved lvalue, then we must reload it */
 974             r = vtop->r;
 975             if ((r & VT_VALMASK) == VT_LLOCAL) {
 976                 SValue v1;
 977                 r = get_reg(RC_INT);
 978                 v1.type.t = VT_INT;
 979                 v1.r = VT_LOCAL | VT_LVAL;
 980                 v1.c.i = fc;
 981                 load(r, &v1);
 982                 fc = 0;
 983             }
 984
 985             if ((ft & VT_BTYPE) == VT_DOUBLE)
 986                 o(0xdc);
 987             else
 988                 o(0xd8);
 989             gen_modrm(a, r, vtop->sym, fc);
 990         }
 991         vtop--;
 992     }
 993 }
 994
 995 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 996    and 'long long' cases. */
 997 ST_FUNC void gen_cvt_itof(int t)
 998 {
 999     save_reg(TREG_ST0);
1000     gv(RC_INT);
1001     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
1002         /* signed long long to float/double/long double (unsigned case
1003            is handled generically) */
1004         o(0x50 + vtop->r2); /* push r2 */
1005         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1006         o(0x242cdf); /* fildll (%esp) */
1007         o(0x08c483); /* add $8, %esp */
1008     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
1009                (VT_INT | VT_UNSIGNED)) {
1010         /* unsigned int to float/double/long double */
1011         o(0x6a); /* push $0 */
1012         g(0x00);
1013         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1014         o(0x242cdf); /* fildll (%esp) */
1015         o(0x08c483); /* add $8, %esp */
1016     } else {
1017         /* int to float/double/long double */
1018         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1019         o(0x2404db); /* fildl (%esp) */
1020         o(0x04c483); /* add $4, %esp */
1021     }
1022     vtop->r = TREG_ST0;
1023 }
1024
1025 /* convert fp to int 't' type */
1026 ST_FUNC void gen_cvt_ftoi(int t)
1027 {
1028  #ifndef COMMIT_4ad186c5ef61_IS_FIXED
1029     /* a good version but it takes a more time to execute */
1030     gv(RC_FLOAT);
1031     save_reg(TREG_EAX);
1032     save_reg(TREG_EDX);
1033     gen_static_call(TOK___tcc_cvt_ftol);
1034     vtop->r = TREG_EAX; /* mark reg as used */
1035     if (t == VT_LLONG)
1036         vtop->r2 = TREG_EDX;
1037  #else
1038     /* a new version with a bug: t2a = 44100312 */
1039     /*
1040     #include<stdio.h>
1041     int main() {
1042         int t1 = 176401255;
1043         float f = 0.25f;
1044         int t2a = (int)(t1 * f); // must be 44100313
1045         int t2b = (int)(t1 * (float)0.25f);
1046         printf("t2a=%d t2b=%d \n",t2a,t2b);
1047         return 0;
1048     }
1049     */
1050     int bt = vtop->type.t & VT_BTYPE;
1051     if (bt == VT_FLOAT)
1052         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1053     else if (bt == VT_LDOUBLE)
1054         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1055     else
1056         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1057     vswap();
1058     gfunc_call(1);
1059     vpushi(0);
1060     vtop->r = REG_IRET;
1061     vtop->r2 = REG_LRET;
1062  #endif
1063 }
1064
1065 /* convert from one floating point type to another */
1066 ST_FUNC void gen_cvt_ftof(int t)
1067 {
1068     /* all we have to do on i386 is to put the float in a register */
1069     gv(RC_FLOAT);
1070 }
1071
1072 /* computed goto support */
1073 ST_FUNC void ggoto(void)
1074 {
1075     gcall_or_jmp(1);
1076     vtop--;
1077 }
1078
1079 /* bound check support functions */
1080 #ifdef CONFIG_TCC_BCHECK
1081
1082 /* generate a bounded pointer addition */
1083 ST_FUNC void gen_bounded_ptr_add(void)
1084 {
1085     /* prepare fast i386 function call (args in eax and edx) */
1086     gv2(RC_EAX, RC_EDX);
1087     /* save all temporary registers */
1088     vtop -= 2;
1089     save_regs(0);
1090     /* do a fast function call */
1091     gen_static_call(TOK___bound_ptr_add);
1092     /* returned pointer is in eax */
1093     vtop++;
1094     vtop->r = TREG_EAX | VT_BOUNDED;
1095     /* address of bounding function call point */
1096     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1097 }
1098
1099 /* patch pointer addition in vtop so that pointer dereferencing is
1100    also tested */
1101 ST_FUNC void gen_bounded_ptr_deref(void)
1102 {
1103     addr_t func;
1104     int  size, align;
1105     Elf32_Rel *rel;
1106     Sym *sym;
1107
1108     size = 0;
1109     /* XXX: put that code in generic part of tcc */
1110     if (!is_float(vtop->type.t)) {
1111         if (vtop->r & VT_LVAL_BYTE)
1112             size = 1;
1113         else if (vtop->r & VT_LVAL_SHORT)
1114             size = 2;
1115     }
1116     if (!size)
1117         size = type_size(&vtop->type, &align);
1118     switch(size) {
1119     case  1: func = TOK___bound_ptr_indir1; break;
1120     case  2: func = TOK___bound_ptr_indir2; break;
1121     case  4: func = TOK___bound_ptr_indir4; break;
1122     case  8: func = TOK___bound_ptr_indir8; break;
1123     case 12: func = TOK___bound_ptr_indir12; break;
1124     case 16: func = TOK___bound_ptr_indir16; break;
1125     default:
1126         tcc_error("unhandled size when dereferencing bounded pointer");
1127         func = 0;
1128         break;
1129     }
1130
1131     /* patch relocation */
1132     /* XXX: find a better solution ? */
1133     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1134     sym = external_global_sym(func, &func_old_type, 0);
1135     if (!sym->c)
1136         put_extern_sym(sym, NULL, 0, 0);
1137     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1138 }
1139 #endif
1140
1141 /* Save the stack pointer onto the stack */
1142 ST_FUNC void gen_vla_sp_save(int addr) {
1143     /* mov %esp,addr(%ebp)*/
1144     o(0x89);
1145     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1146 }
1147
1148 /* Restore the SP from a location on the stack */
1149 ST_FUNC void gen_vla_sp_restore(int addr) {
1150     o(0x8b);
1151     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1152 }
1153
1154 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1155 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1156 #ifdef TCC_TARGET_PE
1157     /* alloca does more than just adjust %rsp on Windows */
1158     vpush_global_sym(&func_old_type, TOK_alloca);
1159     vswap(); /* Move alloca ref past allocation size */
1160     gfunc_call(1);
1161 #else
1162     int r;
1163     r = gv(RC_INT); /* allocation size */
1164     /* sub r,%rsp */
1165     o(0x2b);
1166     o(0xe0 | r);
1167     /* We align to 16 bytes rather than align */
1168     /* and ~15, %esp */
1169     o(0xf0e483);
1170     vpop();
1171 #endif
1172 }
1173
1174 /* end of X86 code generator */
1175 /*************************************************************/
1176 #endif
1177 /*************************************************************/