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 #endif
 111
 112 /* XXX: make it faster ? */
 113 ST_FUNC void g(int c)
 114 {
 115     int ind1;
 116     ind1 = ind + 1;
 117     if (ind1 > cur_text_section->data_allocated)
 118         section_realloc(cur_text_section, ind1);
 119     cur_text_section->data[ind] = c;
 120     ind = ind1;
 121 }
 122
 123 ST_FUNC void o(unsigned int c)
 124 {
 125     while (c) {
 126         g(c);
 127         c = c >> 8;
 128     }
 129 }
 130
 131 ST_FUNC void gen_le16(int v)
 132 {
 133     g(v);
 134     g(v >> 8);
 135 }
 136
 137 ST_FUNC void gen_le32(int c)
 138 {
 139     g(c);
 140     g(c >> 8);
 141     g(c >> 16);
 142     g(c >> 24);
 143 }
 144
 145 /* output a symbol and patch all calls to it */
 146 ST_FUNC void gsym_addr(int t, int a)
 147 {
 148     while (t) {
 149         unsigned char *ptr = cur_text_section->data + t;
 150         uint32_t n = read32le(ptr); /* next value */
 151         write32le(ptr, a - t - 4);
 152         t = n;
 153     }
 154 }
 155
 156 ST_FUNC void gsym(int t)
 157 {
 158     gsym_addr(t, ind);
 159 }
 160
 161 /* psym is used to put an instruction with a data field which is a
 162    reference to a symbol. It is in fact the same as oad ! */
 163 #define psym oad
 164
 165 /* instruction + 4 bytes data. Return the address of the data */
 166 ST_FUNC int oad(int c, int s)
 167 {
 168     int ind1;
 169
 170     o(c);
 171     ind1 = ind + 4;
 172     if (ind1 > cur_text_section->data_allocated)
 173         section_realloc(cur_text_section, ind1);
 174     write32le(cur_text_section->data + ind, s);
 175     s = ind;
 176     ind = ind1;
 177     return s;
 178 }
 179
 180 /* output constant with relocation if 'r & VT_SYM' is true */
 181 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 182 {
 183     if (r & VT_SYM)
 184         greloc(cur_text_section, sym, ind, R_386_32);
 185     gen_le32(c);
 186 }
 187
 188 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 189 {
 190     if (r & VT_SYM)
 191         greloc(cur_text_section, sym, ind, R_386_PC32);
 192     gen_le32(c - 4);
 193 }
 194
 195 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 196    opcode bits */
 197 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 198 {
 199     op_reg = op_reg << 3;
 200     if ((r & VT_VALMASK) == VT_CONST) {
 201         /* constant memory reference */
 202         o(0x05 | op_reg);
 203         gen_addr32(r, sym, c);
 204     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 205         /* currently, we use only ebp as base */
 206         if (c == (char)c) {
 207             /* short reference */
 208             o(0x45 | op_reg);
 209             g(c);
 210         } else {
 211             oad(0x85 | op_reg, c);
 212         }
 213     } else {
 214         g(0x00 | op_reg | (r & VT_VALMASK));
 215     }
 216 }
 217
 218 /* load 'r' from value 'sv' */
 219 ST_FUNC void load(int r, SValue *sv)
 220 {
 221     int v, t, ft, fc, fr;
 222     SValue v1;
 223
 224 #ifdef TCC_TARGET_PE
 225     SValue v2;
 226     sv = pe_getimport(sv, &v2);
 227 #endif
 228
 229     fr = sv->r;
 230     ft = sv->type.t;
 231     fc = sv->c.i;
 232
 233     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 234
 235     v = fr & VT_VALMASK;
 236     if (fr & VT_LVAL) {
 237         if (v == VT_LLOCAL) {
 238             v1.type.t = VT_INT;
 239             v1.r = VT_LOCAL | VT_LVAL;
 240             v1.c.i = fc;
 241             fr = r;
 242             if (!(reg_classes[fr] & RC_INT))
 243                 fr = get_reg(RC_INT);
 244             load(fr, &v1);
 245         }
 246         if ((ft & VT_BTYPE) == VT_FLOAT) {
 247             o(0xd9); /* flds */
 248             r = 0;
 249         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 250             o(0xdd); /* fldl */
 251             r = 0;
 252         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 253             o(0xdb); /* fldt */
 254             r = 5;
 255         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 256             o(0xbe0f);   /* movsbl */
 257         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 258             o(0xb60f);   /* movzbl */
 259         } else if ((ft & VT_TYPE) == VT_SHORT) {
 260             o(0xbf0f);   /* movswl */
 261         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 262             o(0xb70f);   /* movzwl */
 263         } else {
 264             o(0x8b);     /* movl */
 265         }
 266         gen_modrm(r, fr, sv->sym, fc);
 267     } else {
 268         if (v == VT_CONST) {
 269             o(0xb8 + r); /* mov $xx, r */
 270             gen_addr32(fr, sv->sym, fc);
 271         } else if (v == VT_LOCAL) {
 272             if (fc) {
 273                 o(0x8d); /* lea xxx(%ebp), r */
 274                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 275             } else {
 276                 o(0x89);
 277                 o(0xe8 + r); /* mov %ebp, r */
 278             }
 279         } else if (v == VT_CMP) {
 280             oad(0xb8 + r, 0); /* mov $0, r */
 281             o(0x0f); /* setxx %br */
 282             o(fc);
 283             o(0xc0 + r);
 284         } else if (v == VT_JMP || v == VT_JMPI) {
 285             t = v & 1;
 286             oad(0xb8 + r, t); /* mov $1, r */
 287             o(0x05eb); /* jmp after */
 288             gsym(fc);
 289             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 290         } else if (v != r) {
 291             o(0x89);
 292             o(0xc0 + r + v * 8); /* mov v, r */
 293         }
 294     }
 295 }
 296
 297 /* store register 'r' in lvalue 'v' */
 298 ST_FUNC void store(int r, SValue *v)
 299 {
 300     int fr, bt, ft, fc;
 301
 302 #ifdef TCC_TARGET_PE
 303     SValue v2;
 304     v = pe_getimport(v, &v2);
 305 #endif
 306
 307     ft = v->type.t;
 308     fc = v->c.i;
 309     fr = v->r & VT_VALMASK;
 310     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 311     bt = ft & VT_BTYPE;
 312     /* XXX: incorrect if float reg to reg */
 313     if (bt == VT_FLOAT) {
 314         o(0xd9); /* fsts */
 315         r = 2;
 316     } else if (bt == VT_DOUBLE) {
 317         o(0xdd); /* fstpl */
 318         r = 2;
 319     } else if (bt == VT_LDOUBLE) {
 320         o(0xc0d9); /* fld %st(0) */
 321         o(0xdb); /* fstpt */
 322         r = 7;
 323     } else {
 324         if (bt == VT_SHORT)
 325             o(0x66);
 326         if (bt == VT_BYTE || bt == VT_BOOL)
 327             o(0x88);
 328         else
 329             o(0x89);
 330     }
 331     if (fr == VT_CONST ||
 332         fr == VT_LOCAL ||
 333         (v->r & VT_LVAL)) {
 334         gen_modrm(r, v->r, v->sym, fc);
 335     } else if (fr != r) {
 336         o(0xc0 + fr + r * 8); /* mov r, fr */
 337     }
 338 }
 339
 340 static void gadd_sp(int val)
 341 {
 342     if (val == (char)val) {
 343         o(0xc483);
 344         g(val);
 345     } else {
 346         oad(0xc481, val); /* add $xxx, %esp */
 347     }
 348 }
 349
 350 static void gen_static_call(int v)
 351 {
 352     Sym *sym;
 353
 354     sym = external_global_sym(v, &func_old_type, 0);
 355     oad(0xe8, -4);
 356     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 357 }
 358
 359 /* 'is_jmp' is '1' if it is a jump */
 360 static void gcall_or_jmp(int is_jmp)
 361 {
 362     int r;
 363     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 364         int rt;
 365         /* constant case */
 366         if (vtop->r & VT_SYM) {
 367             /* relocation case */
 368             greloc(cur_text_section, vtop->sym,
 369                    ind + 1, R_386_PC32);
 370         } else {
 371             /* put an empty PC32 relocation */
 372             put_elf_reloc(symtab_section, cur_text_section,
 373                           ind + 1, R_386_PC32, 0);
 374         }
 375         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 376         /* extend the return value to the whole register if necessary
 377            visual studio and gcc do not always set the whole eax register
 378            when assigning the return value of a function  */
 379         rt = vtop->type.ref->type.t;
 380         switch (rt & VT_BTYPE) {
 381             case VT_BYTE:
 382                 if (rt & VT_UNSIGNED) {
 383                     o(0xc0b60f); /* movzx %al, %eax */
 384                 }
 385                 else {
 386                     o(0xc0be0f); /* movsx %al, %eax */
 387                 }
 388                 break;
 389             case VT_SHORT:
 390                 if (rt & VT_UNSIGNED) {
 391                     o(0xc0b70f); /* movzx %ax, %eax */
 392                 }
 393                 else {
 394                     o(0xc0bf0f); /* movsx %ax, %eax */
 395                 }
 396                 break;
 397             default:
 398                 break;
 399         }
 400     } else {
 401         /* otherwise, indirect call */
 402         r = gv(RC_INT);
 403         o(0xff); /* call/jmp *r */
 404         o(0xd0 + r + (is_jmp << 4));
 405     }
 406 }
 407
 408 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 409 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 410
 411 /* Return the number of registers needed to return the struct, or 0 if
 412    returning via struct pointer. */
 413 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 414 {
 415 #ifdef TCC_TARGET_PE
 416     int size, align;
 417
 418     *ret_align = 1; // Never have to re-align return values for x86
 419     *regsize = 4;
 420     size = type_size(vt, &align);
 421     if (size > 8) {
 422         return 0;
 423     } else if (size > 4) {
 424         ret->ref = NULL;
 425         ret->t = VT_LLONG;
 426         return 1;
 427     } else {
 428         ret->ref = NULL;
 429         ret->t = VT_INT;
 430         return 1;
 431     }
 432 #else
 433     *ret_align = 1; // Never have to re-align return values for x86
 434     return 0;
 435 #endif
 436 }
 437
 438 /* Generate function call. The function address is pushed first, then
 439    all the parameters in call order. This functions pops all the
 440    parameters and the function address. */
 441 ST_FUNC void gfunc_call(int nb_args)
 442 {
 443     int size, align, r, args_size, i, func_call;
 444     Sym *func_sym;
 445
 446     args_size = 0;
 447     for(i = 0;i < nb_args; i++) {
 448         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 449             size = type_size(&vtop->type, &align);
 450             /* align to stack align size */
 451             size = (size + 3) & ~3;
 452             /* allocate the necessary size on stack */
 453             oad(0xec81, size); /* sub $xxx, %esp */
 454             /* generate structure store */
 455             r = get_reg(RC_INT);
 456             o(0x89); /* mov %esp, r */
 457             o(0xe0 + r);
 458             vset(&vtop->type, r | VT_LVAL, 0);
 459             vswap();
 460             vstore();
 461             args_size += size;
 462         } else if (is_float(vtop->type.t)) {
 463             gv(RC_FLOAT); /* only one float register */
 464             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 465                 size = 4;
 466             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 467                 size = 8;
 468             else
 469                 size = 12;
 470             oad(0xec81, size); /* sub $xxx, %esp */
 471             if (size == 12)
 472                 o(0x7cdb);
 473             else
 474                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 475             g(0x24);
 476             g(0x00);
 477             args_size += size;
 478         } else {
 479             /* simple type (currently always same size) */
 480             /* XXX: implicit cast ? */
 481             r = gv(RC_INT);
 482             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 483                 size = 8;
 484                 o(0x50 + vtop->r2); /* push r */
 485             } else {
 486                 size = 4;
 487             }
 488             o(0x50 + r); /* push r */
 489             args_size += size;
 490         }
 491         vtop--;
 492     }
 493     save_regs(0); /* save used temporary registers */
 494     func_sym = vtop->type.ref;
 495     func_call = func_sym->a.func_call;
 496     /* fast call case */
 497     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 498         func_call == FUNC_FASTCALLW) {
 499         int fastcall_nb_regs;
 500         uint8_t *fastcall_regs_ptr;
 501         if (func_call == FUNC_FASTCALLW) {
 502             fastcall_regs_ptr = fastcallw_regs;
 503             fastcall_nb_regs = 2;
 504         } else {
 505             fastcall_regs_ptr = fastcall_regs;
 506             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 507         }
 508         for(i = 0;i < fastcall_nb_regs; i++) {
 509             if (args_size <= 0)
 510                 break;
 511             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 512             /* XXX: incorrect for struct/floats */
 513             args_size -= 4;
 514         }
 515     }
 516 #ifndef TCC_TARGET_PE
 517     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 518         args_size -= 4;
 519 #endif
 520     gcall_or_jmp(0);
 521
 522     if (args_size && func_call != FUNC_STDCALL)
 523         gadd_sp(args_size);
 524     vtop--;
 525 }
 526
 527 #ifdef TCC_TARGET_PE
 528 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 529 #else
 530 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 531 #endif
 532
 533 /* generate function prolog of type 't' */
 534 ST_FUNC void gfunc_prolog(CType *func_type)
 535 {
 536     int addr, align, size, func_call, fastcall_nb_regs;
 537     int param_index, param_addr;
 538     uint8_t *fastcall_regs_ptr;
 539     Sym *sym;
 540     CType *type;
 541
 542     sym = func_type->ref;
 543     func_call = sym->a.func_call;
 544     addr = 8;
 545     loc = 0;
 546     func_vc = 0;
 547
 548     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 549         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 550         fastcall_regs_ptr = fastcall_regs;
 551     } else if (func_call == FUNC_FASTCALLW) {
 552         fastcall_nb_regs = 2;
 553         fastcall_regs_ptr = fastcallw_regs;
 554     } else {
 555         fastcall_nb_regs = 0;
 556         fastcall_regs_ptr = NULL;
 557     }
 558     param_index = 0;
 559
 560     ind += FUNC_PROLOG_SIZE;
 561     func_sub_sp_offset = ind;
 562     /* if the function returns a structure, then add an
 563        implicit pointer parameter */
 564     func_vt = sym->type;
 565     func_var = (sym->c == FUNC_ELLIPSIS);
 566 #ifdef TCC_TARGET_PE
 567     size = type_size(&func_vt,&align);
 568     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 569 #else
 570     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 571 #endif
 572         /* XXX: fastcall case ? */
 573         func_vc = addr;
 574         addr += 4;
 575         param_index++;
 576     }
 577     /* define parameters */
 578     while ((sym = sym->next) != NULL) {
 579         type = &sym->type;
 580         size = type_size(type, &align);
 581         size = (size + 3) & ~3;
 582 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 583         /* structs are passed as pointer */
 584         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 585             size = 4;
 586         }
 587 #endif
 588         if (param_index < fastcall_nb_regs) {
 589             /* save FASTCALL register */
 590             loc -= 4;
 591             o(0x89);     /* movl */
 592             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 593             param_addr = loc;
 594         } else {
 595             param_addr = addr;
 596             addr += size;
 597         }
 598         sym_push(sym->v & ~SYM_FIELD, type,
 599                  VT_LOCAL | lvalue_type(type->t), param_addr);
 600         param_index++;
 601     }
 602     func_ret_sub = 0;
 603     /* pascal type call ? */
 604     if (func_call == FUNC_STDCALL)
 605         func_ret_sub = addr - 8;
 606 #ifndef TCC_TARGET_PE
 607     else if (func_vc)
 608         func_ret_sub = 4;
 609 #endif
 610
 611 #ifdef CONFIG_TCC_BCHECK
 612     /* leave some room for bound checking code */
 613     if (tcc_state->do_bounds_check) {
 614         oad(0xb8, 0); /* lbound section pointer */
 615         oad(0xb8, 0); /* call to function */
 616         func_bound_offset = lbounds_section->data_offset;
 617     }
 618 #endif
 619 }
 620
 621 /* generate function epilog */
 622 ST_FUNC void gfunc_epilog(void)
 623 {
 624     addr_t v, saved_ind;
 625
 626 #ifdef CONFIG_TCC_BCHECK
 627     if (tcc_state->do_bounds_check
 628      && func_bound_offset != lbounds_section->data_offset) {
 629         addr_t saved_ind;
 630         addr_t *bounds_ptr;
 631         Sym *sym_data;
 632         /* add end of table info */
 633         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 634         *bounds_ptr = 0;
 635         /* generate bound local allocation */
 636         saved_ind = ind;
 637         ind = func_sub_sp_offset;
 638         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 639                                func_bound_offset, lbounds_section->data_offset);
 640         greloc(cur_text_section, sym_data,
 641                ind + 1, R_386_32);
 642         oad(0xb8, 0); /* mov %eax, xxx */
 643         gen_static_call(TOK___bound_local_new);
 644
 645         ind = saved_ind;
 646         /* generate bound check local freeing */
 647         o(0x5250); /* save returned value, if any */
 648         greloc(cur_text_section, sym_data,
 649                ind + 1, R_386_32);
 650         oad(0xb8, 0); /* mov %eax, xxx */
 651         gen_static_call(TOK___bound_local_delete);
 652
 653         o(0x585a); /* restore returned value, if any */
 654     }
 655 #endif
 656     o(0x5b * USE_EBX); /* pop ebx */
 657     o(0xc9); /* leave */
 658     if (func_ret_sub == 0) {
 659         o(0xc3); /* ret */
 660     } else {
 661         o(0xc2); /* ret n */
 662         g(func_ret_sub);
 663         g(func_ret_sub >> 8);
 664     }
 665     /* align local size to word & save local variables */
 666
 667     v = (-loc + 3) & -4;
 668     saved_ind = ind;
 669     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 670 #ifdef TCC_TARGET_PE
 671     if (v >= 4096) {
 672         oad(0xb8, v); /* mov stacksize, %eax */
 673         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 674     } else
 675 #endif
 676     {
 677         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 678         o(0xec81);  /* sub esp, stacksize */
 679         gen_le32(v);
 680 #ifdef TCC_TARGET_PE
 681         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 682 #endif
 683     }
 684     o(0x53 * USE_EBX); /* push ebx */
 685     ind = saved_ind;
 686 }
 687
 688 /* generate a jump to a label */
 689 ST_FUNC int gjmp(int t)
 690 {
 691     return psym(0xe9, t);
 692 }
 693
 694 /* generate a jump to a fixed address */
 695 ST_FUNC void gjmp_addr(int a)
 696 {
 697     int r;
 698     r = a - ind - 2;
 699     if (r == (char)r) {
 700         g(0xeb);
 701         g(r);
 702     } else {
 703         oad(0xe9, a - ind - 5);
 704     }
 705 }
 706
 707 ST_FUNC void gtst_addr(int inv, int a)
 708 {
 709     inv ^= (vtop--)->c.i;
 710     a -= ind + 2;
 711     if (a == (char)a) {
 712         g(inv - 32);
 713         g(a);
 714     } else {
 715         g(0x0f);
 716         oad(inv - 16, a - 4);
 717     }
 718 }
 719
 720 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 721 ST_FUNC int gtst(int inv, int t)
 722 {
 723     int v = vtop->r & VT_VALMASK;
 724     if (v == VT_CMP) {
 725         /* fast case : can jump directly since flags are set */
 726         g(0x0f);
 727         t = psym((vtop->c.i - 16) ^ inv, t);
 728     } else if (v == VT_JMP || v == VT_JMPI) {
 729         /* && or || optimization */
 730         if ((v & 1) == inv) {
 731             /* insert vtop->c jump list in t */
 732             uint32_t n1, n = vtop->c.i;
 733             if (n) {
 734                 while ((n1 = read32le(cur_text_section->data + n)))
 735                     n = n1;
 736                 write32le(cur_text_section->data + n, t);
 737                 t = vtop->c.i;
 738             }
 739         } else {
 740             t = gjmp(t);
 741             gsym(vtop->c.i);
 742         }
 743     }
 744     vtop--;
 745     return t;
 746 }
 747
 748 /* generate an integer binary operation */
 749 ST_FUNC void gen_opi(int op)
 750 {
 751     int r, fr, opc, c;
 752
 753     switch(op) {
 754     case '+':
 755     case TOK_ADDC1: /* add with carry generation */
 756         opc = 0;
 757     gen_op8:
 758         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 759             /* constant case */
 760             vswap();
 761             r = gv(RC_INT);
 762             vswap();
 763             c = vtop->c.i;
 764             if (c == (char)c) {
 765                 /* generate inc and dec for smaller code */
 766                 if (c==1 && opc==0 && op != TOK_ADDC1) {
 767                     o (0x40 | r); // inc
 768                 } else if (c==1 && opc==5 && op != TOK_SUBC1) {
 769                     o (0x48 | r); // dec
 770                 } else {
 771                     o(0x83);
 772                     o(0xc0 | (opc << 3) | r);
 773                     g(c);
 774                 }
 775             } else {
 776                 o(0x81);
 777                 oad(0xc0 | (opc << 3) | r, c);
 778             }
 779         } else {
 780             gv2(RC_INT, RC_INT);
 781             r = vtop[-1].r;
 782             fr = vtop[0].r;
 783             o((opc << 3) | 0x01);
 784             o(0xc0 + r + fr * 8);
 785         }
 786         vtop--;
 787         if (op >= TOK_ULT && op <= TOK_GT) {
 788             vtop->r = VT_CMP;
 789             vtop->c.i = op;
 790         }
 791         break;
 792     case '-':
 793     case TOK_SUBC1: /* sub with carry generation */
 794         opc = 5;
 795         goto gen_op8;
 796     case TOK_ADDC2: /* add with carry use */
 797         opc = 2;
 798         goto gen_op8;
 799     case TOK_SUBC2: /* sub with carry use */
 800         opc = 3;
 801         goto gen_op8;
 802     case '&':
 803         opc = 4;
 804         goto gen_op8;
 805     case '^':
 806         opc = 6;
 807         goto gen_op8;
 808     case '|':
 809         opc = 1;
 810         goto gen_op8;
 811     case '*':
 812         gv2(RC_INT, RC_INT);
 813         r = vtop[-1].r;
 814         fr = vtop[0].r;
 815         vtop--;
 816         o(0xaf0f); /* imul fr, r */
 817         o(0xc0 + fr + r * 8);
 818         break;
 819     case TOK_SHL:
 820         opc = 4;
 821         goto gen_shift;
 822     case TOK_SHR:
 823         opc = 5;
 824         goto gen_shift;
 825     case TOK_SAR:
 826         opc = 7;
 827     gen_shift:
 828         opc = 0xc0 | (opc << 3);
 829         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 830             /* constant case */
 831             vswap();
 832             r = gv(RC_INT);
 833             vswap();
 834             c = vtop->c.i & 0x1f;
 835             o(0xc1); /* shl/shr/sar $xxx, r */
 836             o(opc | r);
 837             g(c);
 838         } else {
 839             /* we generate the shift in ecx */
 840             gv2(RC_INT, RC_ECX);
 841             r = vtop[-1].r;
 842             o(0xd3); /* shl/shr/sar %cl, r */
 843             o(opc | r);
 844         }
 845         vtop--;
 846         break;
 847     case '/':
 848     case TOK_UDIV:
 849     case TOK_PDIV:
 850     case '%':
 851     case TOK_UMOD:
 852     case TOK_UMULL:
 853         /* first operand must be in eax */
 854         /* XXX: need better constraint for second operand */
 855         gv2(RC_EAX, RC_ECX);
 856         r = vtop[-1].r;
 857         fr = vtop[0].r;
 858         vtop--;
 859         save_reg(TREG_EDX);
 860         /* save EAX too if used otherwise */
 861         save_reg_upstack(TREG_EAX, 1);
 862         if (op == TOK_UMULL) {
 863             o(0xf7); /* mul fr */
 864             o(0xe0 + fr);
 865             vtop->r2 = TREG_EDX;
 866             r = TREG_EAX;
 867         } else {
 868             if (op == TOK_UDIV || op == TOK_UMOD) {
 869                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 870                 o(0xf0 + fr);
 871             } else {
 872                 o(0xf799); /* cltd, idiv fr, %eax */
 873                 o(0xf8 + fr);
 874             }
 875             if (op == '%' || op == TOK_UMOD)
 876                 r = TREG_EDX;
 877             else
 878                 r = TREG_EAX;
 879         }
 880         vtop->r = r;
 881         break;
 882     default:
 883         opc = 7;
 884         goto gen_op8;
 885     }
 886 }
 887
 888 /* generate a floating point operation 'v = t1 op t2' instruction. The
 889    two operands are guaranted to have the same floating point type */
 890 /* XXX: need to use ST1 too */
 891 ST_FUNC void gen_opf(int op)
 892 {
 893     int a, ft, fc, swapped, r;
 894
 895     /* convert constants to memory references */
 896     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 897         vswap();
 898         gv(RC_FLOAT);
 899         vswap();
 900     }
 901     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 902         gv(RC_FLOAT);
 903
 904     /* must put at least one value in the floating point register */
 905     if ((vtop[-1].r & VT_LVAL) &&
 906         (vtop[0].r & VT_LVAL)) {
 907         vswap();
 908         gv(RC_FLOAT);
 909         vswap();
 910     }
 911     swapped = 0;
 912     /* swap the stack if needed so that t1 is the register and t2 is
 913        the memory reference */
 914     if (vtop[-1].r & VT_LVAL) {
 915         vswap();
 916         swapped = 1;
 917     }
 918     if (op >= TOK_ULT && op <= TOK_GT) {
 919         /* load on stack second operand */
 920         load(TREG_ST0, vtop);
 921         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 922         if (op == TOK_GE || op == TOK_GT)
 923             swapped = !swapped;
 924         else if (op == TOK_EQ || op == TOK_NE)
 925             swapped = 0;
 926         if (swapped)
 927             o(0xc9d9); /* fxch %st(1) */
 928         if (op == TOK_EQ || op == TOK_NE)
 929             o(0xe9da); /* fucompp */
 930         else
 931             o(0xd9de); /* fcompp */
 932         o(0xe0df); /* fnstsw %ax */
 933         if (op == TOK_EQ) {
 934             o(0x45e480); /* and $0x45, %ah */
 935             o(0x40fC80); /* cmp $0x40, %ah */
 936         } else if (op == TOK_NE) {
 937             o(0x45e480); /* and $0x45, %ah */
 938             o(0x40f480); /* xor $0x40, %ah */
 939             op = TOK_NE;
 940         } else if (op == TOK_GE || op == TOK_LE) {
 941             o(0x05c4f6); /* test $0x05, %ah */
 942             op = TOK_EQ;
 943         } else {
 944             o(0x45c4f6); /* test $0x45, %ah */
 945             op = TOK_EQ;
 946         }
 947         vtop--;
 948         vtop->r = VT_CMP;
 949         vtop->c.i = op;
 950     } else {
 951         /* no memory reference possible for long double operations */
 952         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 953             load(TREG_ST0, vtop);
 954             swapped = !swapped;
 955         }
 956
 957         switch(op) {
 958         default:
 959         case '+':
 960             a = 0;
 961             break;
 962         case '-':
 963             a = 4;
 964             if (swapped)
 965                 a++;
 966             break;
 967         case '*':
 968             a = 1;
 969             break;
 970         case '/':
 971             a = 6;
 972             if (swapped)
 973                 a++;
 974             break;
 975         }
 976         ft = vtop->type.t;
 977         fc = vtop->c.i;
 978         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 979             o(0xde); /* fxxxp %st, %st(1) */
 980             o(0xc1 + (a << 3));
 981         } else {
 982             /* if saved lvalue, then we must reload it */
 983             r = vtop->r;
 984             if ((r & VT_VALMASK) == VT_LLOCAL) {
 985                 SValue v1;
 986                 r = get_reg(RC_INT);
 987                 v1.type.t = VT_INT;
 988                 v1.r = VT_LOCAL | VT_LVAL;
 989                 v1.c.i = fc;
 990                 load(r, &v1);
 991                 fc = 0;
 992             }
 993
 994             if ((ft & VT_BTYPE) == VT_DOUBLE)
 995                 o(0xdc);
 996             else
 997                 o(0xd8);
 998             gen_modrm(a, r, vtop->sym, fc);
 999         }
1000         vtop--;
1001     }
1002 }
1003
1004 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1005    and 'long long' cases. */
1006 ST_FUNC void gen_cvt_itof(int t)
1007 {
1008     save_reg(TREG_ST0);
1009     gv(RC_INT);
1010     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
1011         /* signed long long to float/double/long double (unsigned case
1012            is handled generically) */
1013         o(0x50 + vtop->r2); /* push r2 */
1014         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1015         o(0x242cdf); /* fildll (%esp) */
1016         o(0x08c483); /* add $8, %esp */
1017     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
1018                (VT_INT | VT_UNSIGNED)) {
1019         /* unsigned int to float/double/long double */
1020         o(0x6a); /* push $0 */
1021         g(0x00);
1022         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1023         o(0x242cdf); /* fildll (%esp) */
1024         o(0x08c483); /* add $8, %esp */
1025     } else {
1026         /* int to float/double/long double */
1027         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1028         o(0x2404db); /* fildl (%esp) */
1029         o(0x04c483); /* add $4, %esp */
1030     }
1031     vtop->r = TREG_ST0;
1032 }
1033
1034 /* convert fp to int 't' type */
1035 ST_FUNC void gen_cvt_ftoi(int t)
1036 {
1037 #if 1
1038     gv(RC_FLOAT);
1039     save_reg(TREG_EAX);
1040     save_reg(TREG_EDX);
1041     gen_static_call(TOK___tcc_cvt_ftol);
1042     vtop->r = TREG_EAX; /* mark reg as used */
1043     if (t == VT_LLONG)
1044         vtop->r2 = TREG_EDX;
1045 #else
1046     int bt = vtop->type.t & VT_BTYPE;
1047     if (bt == VT_FLOAT)
1048         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1049     else if (bt == VT_LDOUBLE)
1050         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1051     else
1052         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1053     vswap();
1054     gfunc_call(1);
1055     vpushi(0);
1056     vtop->r = REG_IRET;
1057     vtop->r2 = REG_LRET;
1058 #endif
1059 }
1060
1061 /* convert from one floating point type to another */
1062 ST_FUNC void gen_cvt_ftof(int t)
1063 {
1064     /* all we have to do on i386 is to put the float in a register */
1065     gv(RC_FLOAT);
1066 }
1067
1068 /* computed goto support */
1069 ST_FUNC void ggoto(void)
1070 {
1071     gcall_or_jmp(1);
1072     vtop--;
1073 }
1074
1075 /* bound check support functions */
1076 #ifdef CONFIG_TCC_BCHECK
1077
1078 /* generate a bounded pointer addition */
1079 ST_FUNC void gen_bounded_ptr_add(void)
1080 {
1081     /* prepare fast i386 function call (args in eax and edx) */
1082     gv2(RC_EAX, RC_EDX);
1083     /* save all temporary registers */
1084     vtop -= 2;
1085     save_regs(0);
1086     /* do a fast function call */
1087     gen_static_call(TOK___bound_ptr_add);
1088     /* returned pointer is in eax */
1089     vtop++;
1090     vtop->r = TREG_EAX | VT_BOUNDED;
1091     /* address of bounding function call point */
1092     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1093 }
1094
1095 /* patch pointer addition in vtop so that pointer dereferencing is
1096    also tested */
1097 ST_FUNC void gen_bounded_ptr_deref(void)
1098 {
1099     addr_t func;
1100     int  size, align;
1101     Elf32_Rel *rel;
1102     Sym *sym;
1103
1104     size = 0;
1105     /* XXX: put that code in generic part of tcc */
1106     if (!is_float(vtop->type.t)) {
1107         if (vtop->r & VT_LVAL_BYTE)
1108             size = 1;
1109         else if (vtop->r & VT_LVAL_SHORT)
1110             size = 2;
1111     }
1112     if (!size)
1113         size = type_size(&vtop->type, &align);
1114     switch(size) {
1115     case  1: func = TOK___bound_ptr_indir1; break;
1116     case  2: func = TOK___bound_ptr_indir2; break;
1117     case  4: func = TOK___bound_ptr_indir4; break;
1118     case  8: func = TOK___bound_ptr_indir8; break;
1119     case 12: func = TOK___bound_ptr_indir12; break;
1120     case 16: func = TOK___bound_ptr_indir16; break;
1121     default:
1122         tcc_error("unhandled size when dereferencing bounded pointer");
1123         func = 0;
1124         break;
1125     }
1126
1127     /* patch relocation */
1128     /* XXX: find a better solution ? */
1129     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1130     sym = external_global_sym(func, &func_old_type, 0);
1131     if (!sym->c)
1132         put_extern_sym(sym, NULL, 0, 0);
1133     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1134 }
1135 #endif
1136
1137 /* Save the stack pointer onto the stack */
1138 ST_FUNC void gen_vla_sp_save(int addr) {
1139     /* mov %esp,addr(%ebp)*/
1140     o(0x89);
1141     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1142 }
1143
1144 /* Restore the SP from a location on the stack */
1145 ST_FUNC void gen_vla_sp_restore(int addr) {
1146     o(0x8b);
1147     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1148 }
1149
1150 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1151 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1152 #ifdef TCC_TARGET_PE
1153     /* alloca does more than just adjust %rsp on Windows */
1154     vpush_global_sym(&func_old_type, TOK_alloca);
1155     vswap(); /* Move alloca ref past allocation size */
1156     gfunc_call(1);
1157 #else
1158     int r;
1159     r = gv(RC_INT); /* allocation size */
1160     /* sub r,%rsp */
1161     o(0x2b);
1162     o(0xe0 | r);
1163     /* We align to 16 bytes rather than align */
1164     /* and ~15, %esp */
1165     o(0xf0e483);
1166     vpop();
1167 #endif
1168 }
1169
1170 /* end of X86 code generator */
1171 /*************************************************************/
1172 #endif
1173 /*************************************************************/