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 #define CONFIG_TCC_ASM
  27
  28 /* a register can belong to several classes. The classes must be
  29    sorted from more general to more precise (see gv2() code which does
  30    assumptions on it). */
  31 #define RC_INT     0x0001 /* generic integer register */
  32 #define RC_FLOAT   0x0002 /* generic float register */
  33 #define RC_EAX     0x0004
  34 #define RC_EDX     0x0008
  35 #define RC_ECX     0x0010
  36 #define RC_EBX     0x0020
  37 #define RC_ST0     0x0040
  38
  39 #define RC_IRET    RC_EAX /* function return: integer register */
  40 #define RC_IRE2    RC_EDX /* function return: second integer register */
  41 #define RC_FRET    RC_ST0 /* function return: float register */
  42
  43 /* pretty names for the registers */
  44 enum {
  45     TREG_EAX = 0,
  46     TREG_ECX,
  47     TREG_EDX,
  48     TREG_EBX,
  49     TREG_ST0,
  50     TREG_ESP = 4
  51 };
  52
  53 /* return registers for function */
  54 #define REG_IRET TREG_EAX /* single word int return register */
  55 #define REG_IRE2 TREG_EDX /* second word return register (for long long) */
  56 #define REG_FRET TREG_ST0 /* float return register */
  57
  58 /* defined if function parameters must be evaluated in reverse order */
  59 #define INVERT_FUNC_PARAMS
  60
  61 /* defined if structures are passed as pointers. Otherwise structures
  62    are directly pushed on stack. */
  63 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  64
  65 /* pointer size, in bytes */
  66 #define PTR_SIZE 4
  67
  68 /* long double size and alignment, in bytes */
  69 #define LDOUBLE_SIZE  12
  70 #define LDOUBLE_ALIGN 4
  71 /* maximum alignment (for aligned attribute support) */
  72 #define MAX_ALIGN     8
  73
  74 /* define if return values need to be extended explicitely
  75    at caller side (for interfacing with non-TCC compilers) */
  76 #define PROMOTE_RET
  77
  78 /******************************************************/
  79 #else /* ! TARGET_DEFS_ONLY */
  80 /******************************************************/
  81 #define USING_GLOBALS
  82 #include "tcc.h"
  83
  84 ST_DATA const char * const target_machine_defs =
  85     "__i386__\0"
  86     "__i386\0"
  87     ;
  88
  89 /* define to 1/0 to [not] have EBX as 4th register */
  90 #define USE_EBX 0
  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     /* ebx */ (RC_INT | RC_EBX) * USE_EBX,
  97     /* st0 */ RC_FLOAT | RC_ST0,
  98 };
  99
 100 static unsigned long func_sub_sp_offset;
 101 static int func_ret_sub;
 102 #ifdef CONFIG_TCC_BCHECK
 103 static addr_t func_bound_offset;
 104 static unsigned long func_bound_ind;
 105 ST_DATA int func_bound_add_epilog;
 106 static void gen_bounds_prolog(void);
 107 static void gen_bounds_epilog(void);
 108 #endif
 109
 110 /* XXX: make it faster ? */
 111 ST_FUNC void g(int c)
 112 {
 113     int ind1;
 114     if (nocode_wanted)
 115         return;
 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 /* instruction + 4 bytes data. Return the address of the data */
 157 static int oad(int c, int s)
 158 {
 159     int t;
 160     if (nocode_wanted)
 161         return s;
 162     o(c);
 163     t = ind;
 164     gen_le32(s);
 165     return t;
 166 }
 167
 168 ST_FUNC void gen_fill_nops(int bytes)
 169 {
 170     while (bytes--)
 171       g(0x90);
 172 }
 173
 174 /* generate jmp to a label */
 175 #define gjmp2(instr,lbl) oad(instr,lbl)
 176
 177 /* output constant with relocation if 'r & VT_SYM' is true */
 178 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 179 {
 180     if (r & VT_SYM)
 181         greloc(cur_text_section, sym, ind, R_386_32);
 182     gen_le32(c);
 183 }
 184
 185 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 186 {
 187     if (r & VT_SYM)
 188         greloc(cur_text_section, sym, ind, R_386_PC32);
 189     gen_le32(c - 4);
 190 }
 191
 192 /* generate a modrm reference. 'op_reg' contains the additional 3
 193    opcode bits */
 194 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 195 {
 196     op_reg = op_reg << 3;
 197     if ((r & VT_VALMASK) == VT_CONST) {
 198         /* constant memory reference */
 199         o(0x05 | op_reg);
 200         gen_addr32(r, sym, c);
 201     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 202         /* currently, we use only ebp as base */
 203         if (c == (char)c) {
 204             /* short reference */
 205             o(0x45 | op_reg);
 206             g(c);
 207         } else {
 208             oad(0x85 | op_reg, c);
 209         }
 210     } else {
 211         g(0x00 | op_reg | (r & VT_VALMASK));
 212     }
 213 }
 214
 215 /* load 'r' from value 'sv' */
 216 ST_FUNC void load(int r, SValue *sv)
 217 {
 218     int v, t, ft, fc, fr;
 219     SValue v1;
 220
 221 #ifdef TCC_TARGET_PE
 222     SValue v2;
 223     sv = pe_getimport(sv, &v2);
 224 #endif
 225
 226     fr = sv->r;
 227     ft = sv->type.t & ~VT_DEFSIGN;
 228     fc = sv->c.i;
 229
 230     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 231
 232     v = fr & VT_VALMASK;
 233     if (fr & VT_LVAL) {
 234         if (v == VT_LLOCAL) {
 235             v1.type.t = VT_INT;
 236             v1.r = VT_LOCAL | VT_LVAL;
 237             v1.c.i = fc;
 238             v1.sym = NULL;
 239             fr = r;
 240             if (!(reg_classes[fr] & RC_INT))
 241                 fr = get_reg(RC_INT);
 242             load(fr, &v1);
 243         }
 244         if ((ft & VT_BTYPE) == VT_FLOAT) {
 245             o(0xd9); /* flds */
 246             r = 0;
 247         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 248             o(0xdd); /* fldl */
 249             r = 0;
 250         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 251             o(0xdb); /* fldt */
 252             r = 5;
 253         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 254             o(0xbe0f);   /* movsbl */
 255         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 256             o(0xb60f);   /* movzbl */
 257         } else if ((ft & VT_TYPE) == VT_SHORT) {
 258             o(0xbf0f);   /* movswl */
 259         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 260             o(0xb70f);   /* movzwl */
 261         } else {
 262             o(0x8b);     /* movl */
 263         }
 264         gen_modrm(r, fr, sv->sym, fc);
 265     } else {
 266         if (v == VT_CONST) {
 267             o(0xb8 + r); /* mov $xx, r */
 268             gen_addr32(fr, sv->sym, fc);
 269         } else if (v == VT_LOCAL) {
 270             if (fc) {
 271                 o(0x8d); /* lea xxx(%ebp), r */
 272                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 273             } else {
 274                 o(0x89);
 275                 o(0xe8 + r); /* mov %ebp, r */
 276             }
 277         } else if (v == VT_CMP) {
 278             o(0x0f); /* setxx %br */
 279             o(fc);
 280             o(0xc0 + r);
 281             o(0xc0b60f + r * 0x90000); /* movzbl %al, %eax */
 282         } else if (v == VT_JMP || v == VT_JMPI) {
 283             t = v & 1;
 284             oad(0xb8 + r, t); /* mov $1, r */
 285             o(0x05eb); /* jmp after */
 286             gsym(fc);
 287             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 288         } else if (v != r) {
 289             o(0x89);
 290             o(0xc0 + r + v * 8); /* mov v, r */
 291         }
 292     }
 293 }
 294
 295 /* store register 'r' in lvalue 'v' */
 296 ST_FUNC void store(int r, SValue *v)
 297 {
 298     int fr, bt, ft, fc;
 299
 300 #ifdef TCC_TARGET_PE
 301     SValue v2;
 302     v = pe_getimport(v, &v2);
 303 #endif
 304
 305     ft = v->type.t;
 306     fc = v->c.i;
 307     fr = v->r & VT_VALMASK;
 308     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 309     bt = ft & VT_BTYPE;
 310     /* XXX: incorrect if float reg to reg */
 311     if (bt == VT_FLOAT) {
 312         o(0xd9); /* fsts */
 313         r = 2;
 314     } else if (bt == VT_DOUBLE) {
 315         o(0xdd); /* fstpl */
 316         r = 2;
 317     } else if (bt == VT_LDOUBLE) {
 318         o(0xc0d9); /* fld %st(0) */
 319         o(0xdb); /* fstpt */
 320         r = 7;
 321     } else {
 322         if (bt == VT_SHORT)
 323             o(0x66);
 324         if (bt == VT_BYTE || bt == VT_BOOL)
 325             o(0x88);
 326         else
 327             o(0x89);
 328     }
 329     if (fr == VT_CONST ||
 330         fr == VT_LOCAL ||
 331         (v->r & VT_LVAL)) {
 332         gen_modrm(r, v->r, v->sym, fc);
 333     } else if (fr != r) {
 334         o(0xc0 + fr + r * 8); /* mov r, fr */
 335     }
 336 }
 337
 338 static void gadd_sp(int val)
 339 {
 340     if (val == (char)val) {
 341         o(0xc483);
 342         g(val);
 343     } else {
 344         oad(0xc481, val); /* add $xxx, %esp */
 345     }
 346 }
 347
 348 #if defined CONFIG_TCC_BCHECK || defined TCC_TARGET_PE
 349 static void gen_static_call(int v)
 350 {
 351     Sym *sym;
 352
 353     sym = external_helper_sym(v);
 354     oad(0xe8, -4);
 355     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 356 }
 357 #endif
 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 && (vtop->r & VT_SYM)) {
 364         /* constant and relocation case */
 365         greloc(cur_text_section, vtop->sym, ind + 1, R_386_PC32);
 366         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 367     } else {
 368         /* otherwise, indirect call */
 369         r = gv(RC_INT);
 370         o(0xff); /* call/jmp *r */
 371         o(0xd0 + r + (is_jmp << 4));
 372     }
 373 }
 374
 375 static const uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 376 static const uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 377
 378 /* Return the number of registers needed to return the struct, or 0 if
 379    returning via struct pointer. */
 380 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 381 {
 382 #if defined(TCC_TARGET_PE) || TARGETOS_FreeBSD || TARGETOS_OpenBSD
 383     int size, align, nregs;
 384     *ret_align = 1; // Never have to re-align return values for x86
 385     *regsize = 4;
 386     size = type_size(vt, &align);
 387     if (size > 8 || (size & (size - 1)))
 388         return 0;
 389     nregs = 1;
 390     if (size == 8)
 391         ret->t = VT_INT, nregs = 2;
 392     else if (size == 4)
 393         ret->t = VT_INT;
 394     else if (size == 2)
 395         ret->t = VT_SHORT;
 396     else
 397         ret->t = VT_BYTE;
 398     ret->ref = NULL;
 399     return nregs;
 400 #else
 401     *ret_align = 1; // Never have to re-align return values for x86
 402     return 0;
 403 #endif
 404 }
 405
 406 /* Generate function call. The function address is pushed first, then
 407    all the parameters in call order. This functions pops all the
 408    parameters and the function address. */
 409 ST_FUNC void gfunc_call(int nb_args)
 410 {
 411     int size, align, r, args_size, i, func_call;
 412     Sym *func_sym;
 413     // Look ahead to the function on the stack to get the function call type
 414     int func_call2 = ((vtop - nb_args)->type.ref)->f.func_call;
 415
 416 #ifdef CONFIG_TCC_BCHECK
 417     if (tcc_state->do_bounds_check)
 418         gbound_args(nb_args);
 419 #endif
 420
 421     args_size = 0;
 422     for(i = 0;i < nb_args; i++) {
 423         if (func_call2 == FUNC_THISCALL && i == (nb_args - 1)) {
 424             // If thiscall, zap the last push, as it is `this`. Instead, mov into ecx
 425             size = 0;
 426             load(get_reg(RC_ECX), vtop);
 427         }
 428         else
 429         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 430             size = type_size(&vtop->type, &align);
 431             /* align to stack align size */
 432             size = (size + 3) & ~3;
 433             /* allocate the necessary size on stack */
 434 #ifdef TCC_TARGET_PE
 435             if (size >= 4096) {
 436                 r = get_reg(RC_EAX);
 437                 oad(0x68, size); // push size
 438                 /* cannot call normal 'alloca' with bound checking */
 439                 gen_static_call(tok_alloc_const("__alloca"));
 440                 gadd_sp(4);
 441             } else
 442 #endif
 443             {
 444                 oad(0xec81, size); /* sub $xxx, %esp */
 445                 /* generate structure store */
 446                 r = get_reg(RC_INT);
 447                 o(0xe089 + (r << 8)); /* mov %esp, r */
 448             }
 449             vset(&vtop->type, r | VT_LVAL, 0);
 450             vswap();
 451             vstore();
 452             args_size += size;
 453         } else if (is_float(vtop->type.t)) {
 454             gv(RC_FLOAT); /* only one float register */
 455             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 456                 size = 4;
 457             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 458                 size = 8;
 459             else
 460                 size = 12;
 461             oad(0xec81, size); /* sub $xxx, %esp */
 462             if (size == 12)
 463                 o(0x7cdb);
 464             else
 465                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 466             g(0x24);
 467             g(0x00);
 468             args_size += size;
 469         } else {
 470             /* simple type (currently always same size) */
 471             /* XXX: implicit cast ? */
 472             r = gv(RC_INT);
 473             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 474                 size = 8;
 475                 o(0x50 + vtop->r2); /* push r */
 476             } else {
 477                 size = 4;
 478             }
 479             o(0x50 + r); /* push r */
 480             args_size += size;
 481         }
 482         vtop--;
 483     }
 484     save_regs(0); /* save used temporary registers */
 485     func_sym = vtop->type.ref;
 486     func_call = func_sym->f.func_call;
 487     /* fast call case */
 488     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 489         func_call == FUNC_FASTCALLW) {
 490         int fastcall_nb_regs;
 491         const uint8_t *fastcall_regs_ptr;
 492         if (func_call == FUNC_FASTCALLW) {
 493             fastcall_regs_ptr = fastcallw_regs;
 494             fastcall_nb_regs = 2;
 495         } else {
 496             fastcall_regs_ptr = fastcall_regs;
 497             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 498         }
 499         for(i = 0;i < fastcall_nb_regs; i++) {
 500             if (args_size <= 0)
 501                 break;
 502             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 503             /* XXX: incorrect for struct/floats */
 504             args_size -= 4;
 505         }
 506     }
 507 #if !defined(TCC_TARGET_PE) && !TARGETOS_FreeBSD || TARGETOS_OpenBSD
 508     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 509         args_size -= 4;
 510 #endif
 511
 512     gcall_or_jmp(0);
 513
 514     if (args_size && func_call != FUNC_STDCALL && func_call != FUNC_THISCALL && func_call != FUNC_FASTCALLW)
 515         gadd_sp(args_size);
 516     vtop--;
 517 }
 518
 519 #ifdef TCC_TARGET_PE
 520 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 521 #else
 522 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 523 #endif
 524
 525 /* generate function prolog of type 't' */
 526 ST_FUNC void gfunc_prolog(Sym *func_sym)
 527 {
 528     CType *func_type = &func_sym->type;
 529     int addr, align, size, func_call, fastcall_nb_regs;
 530     int param_index, param_addr;
 531     const uint8_t *fastcall_regs_ptr;
 532     Sym *sym;
 533     CType *type;
 534     int thiscall_nb_regs;
 535
 536     sym = func_type->ref;
 537     func_call = sym->f.func_call;
 538     addr = 8;
 539     loc = 0;
 540     func_vc = 0;
 541
 542     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 543         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 544         fastcall_regs_ptr = fastcall_regs;
 545     } else if (func_call == FUNC_FASTCALLW) {
 546         fastcall_nb_regs = 2;
 547         fastcall_regs_ptr = fastcallw_regs;
 548     } else {
 549         fastcall_nb_regs = 0;
 550         fastcall_regs_ptr = NULL;
 551     }
 552
 553     if (func_call == FUNC_THISCALL) {
 554         thiscall_nb_regs = 1;
 555     }
 556
 557     param_index = 0;
 558
 559     ind += FUNC_PROLOG_SIZE;
 560     func_sub_sp_offset = ind;
 561     /* if the function returns a structure, then add an
 562        implicit pointer parameter */
 563 #if defined(TCC_TARGET_PE) || TARGETOS_FreeBSD || TARGETOS_OpenBSD
 564     size = type_size(&func_vt,&align);
 565     if (((func_vt.t & VT_BTYPE) == VT_STRUCT)
 566         && (size > 8 || (size & (size - 1)))) {
 567 #else
 568     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 569 #endif
 570         /* XXX: fastcall case ? */
 571         func_vc = addr;
 572         addr += 4;
 573         param_index++;
 574     }
 575     /* define parameters */
 576     while ((sym = sym->next) != NULL) {
 577         type = &sym->type;
 578         size = type_size(type, &align);
 579         size = (size + 3) & ~3;
 580 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 581         /* structs are passed as pointer */
 582         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 583             size = 4;
 584         }
 585 #endif
 586         if (param_index < fastcall_nb_regs) {
 587             /* save FASTCALL register */
 588             loc -= 4;
 589             o(0x89);     /* movl */
 590             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 591             param_addr = loc;
 592         }
 593         else if(param_index < thiscall_nb_regs) {
 594             /* Why ? */
 595             /* save THISCALL register; ECX */
 596             loc -= 4;
 597             o(0x89);     /* movl */
 598             gen_modrm(TREG_ECX, VT_LOCAL, NULL, loc);
 599             param_addr = loc;
 600         } else {
 601             param_addr = addr;
 602             addr += size;
 603         }
 604         sym_push(sym->v & ~SYM_FIELD, type,
 605                  VT_LOCAL | VT_LVAL, param_addr);
 606         param_index++;
 607     }
 608     func_ret_sub = 0;
 609     /* pascal type call or fastcall ? */
 610     if (func_call == FUNC_STDCALL || func_call == FUNC_FASTCALLW || func_call == FUNC_THISCALL)
 611         func_ret_sub = addr - 8;
 612 #if !defined(TCC_TARGET_PE) && !TARGETOS_FreeBSD || TARGETOS_OpenBSD
 613     else if (func_vc)
 614         func_ret_sub = 4;
 615 #endif
 616
 617 #ifdef CONFIG_TCC_BCHECK
 618     if (tcc_state->do_bounds_check)
 619         gen_bounds_prolog();
 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         gen_bounds_epilog();
 631 #endif
 632
 633     /* align local size to word & save local variables */
 634     v = (-loc + 3) & -4;
 635
 636 #if USE_EBX
 637     o(0x8b);
 638     gen_modrm(TREG_EBX, VT_LOCAL, NULL, -(v+4));
 639 #endif
 640
 641     o(0xc9); /* leave */
 642     if (func_ret_sub == 0) {
 643         o(0xc3); /* ret */
 644     } else {
 645         o(0xc2); /* ret n */
 646         g(func_ret_sub);
 647         g(func_ret_sub >> 8);
 648     }
 649     saved_ind = ind;
 650     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 651 #ifdef TCC_TARGET_PE
 652     if (v >= 4096) {
 653         oad(0xb8, v); /* mov stacksize, %eax */
 654         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 655     } else
 656 #endif
 657     {
 658         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 659         o(0xec81);  /* sub esp, stacksize */
 660         gen_le32(v);
 661 #ifdef TCC_TARGET_PE
 662         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 663 #endif
 664     }
 665     o(0x53 * USE_EBX); /* push ebx */
 666     ind = saved_ind;
 667 }
 668
 669 /* generate a jump to a label */
 670 ST_FUNC int gjmp(int t)
 671 {
 672     return gjmp2(0xe9, t);
 673 }
 674
 675 /* generate a jump to a fixed address */
 676 ST_FUNC void gjmp_addr(int a)
 677 {
 678     int r;
 679     r = a - ind - 2;
 680     if (r == (char)r) {
 681         g(0xeb);
 682         g(r);
 683     } else {
 684         oad(0xe9, a - ind - 5);
 685     }
 686 }
 687
 688 #if 0
 689 /* generate a jump to a fixed address */
 690 ST_FUNC void gjmp_cond_addr(int a, int op)
 691 {
 692     int r = a - ind - 2;
 693     if (r == (char)r)
 694         g(op - 32), g(r);
 695     else
 696         g(0x0f), gjmp2(op - 16, r - 4);
 697 }
 698 #endif
 699
 700 ST_FUNC int gjmp_append(int n, int t)
 701 {
 702     void *p;
 703     /* insert vtop->c jump list in t */
 704     if (n) {
 705         uint32_t n1 = n, n2;
 706         while ((n2 = read32le(p = cur_text_section->data + n1)))
 707             n1 = n2;
 708         write32le(p, t);
 709         t = n;
 710     }
 711     return t;
 712 }
 713
 714 ST_FUNC int gjmp_cond(int op, int t)
 715 {
 716     g(0x0f);
 717     t = gjmp2(op - 16, t);
 718     return t;
 719 }
 720
 721 ST_FUNC void gen_opi(int op)
 722 {
 723     int r, fr, opc, c;
 724
 725     switch(op) {
 726     case '+':
 727     case TOK_ADDC1: /* add with carry generation */
 728         opc = 0;
 729     gen_op8:
 730         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 731             /* constant case */
 732             vswap();
 733             r = gv(RC_INT);
 734             vswap();
 735             c = vtop->c.i;
 736             if (c == (char)c) {
 737                 /* generate inc and dec for smaller code */
 738                 if ((c == 1 || c == -1) && (op == '+' || op == '-')) {
 739                     opc = (c == 1) ^ (op == '+');
 740                     o (0x40 | (opc << 3) | r); // inc,dec
 741                 } else {
 742                     o(0x83);
 743                     o(0xc0 | (opc << 3) | r);
 744                     g(c);
 745                 }
 746             } else {
 747                 o(0x81);
 748                 oad(0xc0 | (opc << 3) | r, c);
 749             }
 750         } else {
 751             gv2(RC_INT, RC_INT);
 752             r = vtop[-1].r;
 753             fr = vtop[0].r;
 754             o((opc << 3) | 0x01);
 755             o(0xc0 + r + fr * 8);
 756         }
 757         vtop--;
 758         if (op >= TOK_ULT && op <= TOK_GT)
 759             vset_VT_CMP(op);
 760         break;
 761     case '-':
 762     case TOK_SUBC1: /* sub with carry generation */
 763         opc = 5;
 764         goto gen_op8;
 765     case TOK_ADDC2: /* add with carry use */
 766         opc = 2;
 767         goto gen_op8;
 768     case TOK_SUBC2: /* sub with carry use */
 769         opc = 3;
 770         goto gen_op8;
 771     case '&':
 772         opc = 4;
 773         goto gen_op8;
 774     case '^':
 775         opc = 6;
 776         goto gen_op8;
 777     case '|':
 778         opc = 1;
 779         goto gen_op8;
 780     case '*':
 781         gv2(RC_INT, RC_INT);
 782         r = vtop[-1].r;
 783         fr = vtop[0].r;
 784         vtop--;
 785         o(0xaf0f); /* imul fr, r */
 786         o(0xc0 + fr + r * 8);
 787         break;
 788     case TOK_SHL:
 789         opc = 4;
 790         goto gen_shift;
 791     case TOK_SHR:
 792         opc = 5;
 793         goto gen_shift;
 794     case TOK_SAR:
 795         opc = 7;
 796     gen_shift:
 797         opc = 0xc0 | (opc << 3);
 798         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 799             /* constant case */
 800             vswap();
 801             r = gv(RC_INT);
 802             vswap();
 803             c = vtop->c.i & 0x1f;
 804             o(0xc1); /* shl/shr/sar $xxx, r */
 805             o(opc | r);
 806             g(c);
 807         } else {
 808             /* we generate the shift in ecx */
 809             gv2(RC_INT, RC_ECX);
 810             r = vtop[-1].r;
 811             o(0xd3); /* shl/shr/sar %cl, r */
 812             o(opc | r);
 813         }
 814         vtop--;
 815         break;
 816     case '/':
 817     case TOK_UDIV:
 818     case TOK_PDIV:
 819     case '%':
 820     case TOK_UMOD:
 821     case TOK_UMULL:
 822         /* first operand must be in eax */
 823         /* XXX: need better constraint for second operand */
 824         gv2(RC_EAX, RC_ECX);
 825         r = vtop[-1].r;
 826         fr = vtop[0].r;
 827         vtop--;
 828         save_reg(TREG_EDX);
 829         /* save EAX too if used otherwise */
 830         save_reg_upstack(TREG_EAX, 1);
 831         if (op == TOK_UMULL) {
 832             o(0xf7); /* mul fr */
 833             o(0xe0 + fr);
 834             vtop->r2 = TREG_EDX;
 835             r = TREG_EAX;
 836         } else {
 837             if (op == TOK_UDIV || op == TOK_UMOD) {
 838                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 839                 o(0xf0 + fr);
 840             } else {
 841                 o(0xf799); /* cltd, idiv fr, %eax */
 842                 o(0xf8 + fr);
 843             }
 844             if (op == '%' || op == TOK_UMOD)
 845                 r = TREG_EDX;
 846             else
 847                 r = TREG_EAX;
 848         }
 849         vtop->r = r;
 850         break;
 851     default:
 852         opc = 7;
 853         goto gen_op8;
 854     }
 855 }
 856
 857 /* generate a floating point operation 'v = t1 op t2' instruction. The
 858    two operands are guaranteed to have the same floating point type */
 859 /* XXX: need to use ST1 too */
 860 ST_FUNC void gen_opf(int op)
 861 {
 862     int a, ft, fc, swapped, r;
 863
 864     if (op == TOK_NEG) { /* unary minus */
 865         gv(RC_FLOAT);
 866         o(0xe0d9); /* fchs */
 867         return;
 868     }
 869
 870     /* convert constants to memory references */
 871     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 872         vswap();
 873         gv(RC_FLOAT);
 874         vswap();
 875     }
 876     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 877         gv(RC_FLOAT);
 878
 879     /* must put at least one value in the floating point register */
 880     if ((vtop[-1].r & VT_LVAL) &&
 881         (vtop[0].r & VT_LVAL)) {
 882         vswap();
 883         gv(RC_FLOAT);
 884         vswap();
 885     }
 886     swapped = 0;
 887     /* swap the stack if needed so that t1 is the register and t2 is
 888        the memory reference */
 889     if (vtop[-1].r & VT_LVAL) {
 890         vswap();
 891         swapped = 1;
 892     }
 893     if (op >= TOK_ULT && op <= TOK_GT) {
 894         /* load on stack second operand */
 895         load(TREG_ST0, vtop);
 896         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 897         if (op == TOK_GE || op == TOK_GT)
 898             swapped = !swapped;
 899         else if (op == TOK_EQ || op == TOK_NE)
 900             swapped = 0;
 901         if (swapped)
 902             o(0xc9d9); /* fxch %st(1) */
 903         if (op == TOK_EQ || op == TOK_NE)
 904             o(0xe9da); /* fucompp */
 905         else
 906             o(0xd9de); /* fcompp */
 907         o(0xe0df); /* fnstsw %ax */
 908         if (op == TOK_EQ) {
 909             o(0x45e480); /* and $0x45, %ah */
 910             o(0x40fC80); /* cmp $0x40, %ah */
 911         } else if (op == TOK_NE) {
 912             o(0x45e480); /* and $0x45, %ah */
 913             o(0x40f480); /* xor $0x40, %ah */
 914             op = TOK_NE;
 915         } else if (op == TOK_GE || op == TOK_LE) {
 916             o(0x05c4f6); /* test $0x05, %ah */
 917             op = TOK_EQ;
 918         } else {
 919             o(0x45c4f6); /* test $0x45, %ah */
 920             op = TOK_EQ;
 921         }
 922         vtop--;
 923         vset_VT_CMP(op);
 924     } else {
 925         /* no memory reference possible for long double operations */
 926         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 927             load(TREG_ST0, vtop);
 928             swapped = !swapped;
 929         }
 930
 931         switch(op) {
 932         default:
 933         case '+':
 934             a = 0;
 935             break;
 936         case '-':
 937             a = 4;
 938             if (swapped)
 939                 a++;
 940             break;
 941         case '*':
 942             a = 1;
 943             break;
 944         case '/':
 945             a = 6;
 946             if (swapped)
 947                 a++;
 948             break;
 949         }
 950         ft = vtop->type.t;
 951         fc = vtop->c.i;
 952         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 953             o(0xde); /* fxxxp %st, %st(1) */
 954             o(0xc1 + (a << 3));
 955         } else {
 956             /* if saved lvalue, then we must reload it */
 957             r = vtop->r;
 958             if ((r & VT_VALMASK) == VT_LLOCAL) {
 959                 SValue v1;
 960                 r = get_reg(RC_INT);
 961                 v1.type.t = VT_INT;
 962                 v1.r = VT_LOCAL | VT_LVAL;
 963                 v1.c.i = fc;
 964                 v1.sym = NULL;
 965                 load(r, &v1);
 966                 fc = 0;
 967             }
 968
 969             if ((ft & VT_BTYPE) == VT_DOUBLE)
 970                 o(0xdc);
 971             else
 972                 o(0xd8);
 973             gen_modrm(a, r, vtop->sym, fc);
 974         }
 975         vtop--;
 976     }
 977 }
 978
 979 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 980    and 'long long' cases. */
 981 ST_FUNC void gen_cvt_itof(int t)
 982 {
 983     save_reg(TREG_ST0);
 984     gv(RC_INT);
 985     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 986         /* signed long long to float/double/long double (unsigned case
 987            is handled generically) */
 988         o(0x50 + vtop->r2); /* push r2 */
 989         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 990         o(0x242cdf); /* fildll (%esp) */
 991         o(0x08c483); /* add $8, %esp */
 992         vtop->r2 = VT_CONST;
 993     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 994                (VT_INT | VT_UNSIGNED)) {
 995         /* unsigned int to float/double/long double */
 996         o(0x6a); /* push $0 */
 997         g(0x00);
 998         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 999         o(0x242cdf); /* fildll (%esp) */
1000         o(0x08c483); /* add $8, %esp */
1001     } else {
1002         /* int to float/double/long double */
1003         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1004         o(0x2404db); /* fildl (%esp) */
1005         o(0x04c483); /* add $4, %esp */
1006     }
1007     vtop->r2 = VT_CONST;
1008     vtop->r = TREG_ST0;
1009 }
1010
1011 /* convert fp to int 't' type */
1012 ST_FUNC void gen_cvt_ftoi(int t)
1013 {
1014     int bt = vtop->type.t & VT_BTYPE;
1015     if (bt == VT_FLOAT)
1016         vpush_helper_func(TOK___fixsfdi);
1017     else if (bt == VT_LDOUBLE)
1018         vpush_helper_func(TOK___fixxfdi);
1019     else
1020         vpush_helper_func(TOK___fixdfdi);
1021     vswap();
1022     gfunc_call(1);
1023     vpushi(0);
1024     vtop->r = REG_IRET;
1025     if ((t & VT_BTYPE) == VT_LLONG)
1026         vtop->r2 = REG_IRE2;
1027 }
1028
1029 /* convert from one floating point type to another */
1030 ST_FUNC void gen_cvt_ftof(int t)
1031 {
1032     /* all we have to do on i386 is to put the float in a register */
1033     gv(RC_FLOAT);
1034 }
1035
1036 /* char/short to int conversion */
1037 ST_FUNC void gen_cvt_csti(int t)
1038 {
1039     int r, sz, xl;
1040     r = gv(RC_INT);
1041     sz = !(t & VT_UNSIGNED);
1042     xl = (t & VT_BTYPE) == VT_SHORT;
1043     o(0xc0b60f /* mov[sz] %a[xl], %eax */
1044         | (sz << 3 | xl) << 8
1045         | (r << 3 | r) << 16
1046         );
1047 }
1048
1049 /* increment tcov counter */
1050 ST_FUNC void gen_increment_tcov (SValue *sv)
1051 {
1052    o(0x0583); /* addl $1, xxx */
1053    greloc(cur_text_section, sv->sym, ind, R_386_32);
1054    gen_le32(0);
1055    o(1);
1056    o(0x1583); /* addcl $0, xxx */
1057    greloc(cur_text_section, sv->sym, ind, R_386_32);
1058    gen_le32(4);
1059    g(0);
1060 }
1061
1062 /* computed goto support */
1063 ST_FUNC void ggoto(void)
1064 {
1065     gcall_or_jmp(1);
1066     vtop--;
1067 }
1068
1069 /* bound check support functions */
1070 #ifdef CONFIG_TCC_BCHECK
1071
1072 static void gen_bounds_prolog(void)
1073 {
1074     /* leave some room for bound checking code */
1075     func_bound_offset = lbounds_section->data_offset;
1076     func_bound_ind = ind;
1077     func_bound_add_epilog = 0;
1078     oad(0xb8, 0); /* lbound section pointer */
1079     oad(0xb8, 0); /* call to function */
1080 }
1081
1082 static void gen_bounds_epilog(void)
1083 {
1084     addr_t saved_ind;
1085     addr_t *bounds_ptr;
1086     Sym *sym_data;
1087     int offset_modified = func_bound_offset != lbounds_section->data_offset;
1088
1089     if (!offset_modified && !func_bound_add_epilog)
1090         return;
1091
1092     /* add end of table info */
1093     bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
1094     *bounds_ptr = 0;
1095
1096     sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
1097                            func_bound_offset, PTR_SIZE);
1098
1099     /* generate bound local allocation */
1100     if (offset_modified) {
1101         saved_ind = ind;
1102         ind = func_bound_ind;
1103         greloc(cur_text_section, sym_data, ind + 1, R_386_32);
1104         ind = ind + 5;
1105         gen_static_call(TOK___bound_local_new);
1106         ind = saved_ind;
1107     }
1108
1109     /* generate bound check local freeing */
1110     o(0x5250); /* save returned value, if any */
1111     greloc(cur_text_section, sym_data, ind + 1, R_386_32);
1112     oad(0xb8, 0); /* mov %eax, xxx */
1113     gen_static_call(TOK___bound_local_delete);
1114     o(0x585a); /* restore returned value, if any */
1115 }
1116 #endif
1117
1118 /* Save the stack pointer onto the stack */
1119 ST_FUNC void gen_vla_sp_save(int addr) {
1120     /* mov %esp,addr(%ebp)*/
1121     o(0x89);
1122     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1123 }
1124
1125 /* Restore the SP from a location on the stack */
1126 ST_FUNC void gen_vla_sp_restore(int addr) {
1127     o(0x8b);
1128     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1129 }
1130
1131 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1132 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1133     int use_call = 0;
1134
1135 #if defined(CONFIG_TCC_BCHECK)
1136     use_call = tcc_state->do_bounds_check;
1137 #endif
1138 #ifdef TCC_TARGET_PE    /* alloca does more than just adjust %rsp on Windows */
1139     use_call = 1;
1140 #endif
1141     if (use_call)
1142     {
1143         vpush_helper_func(TOK_alloca);
1144         vswap(); /* Move alloca ref past allocation size */
1145         gfunc_call(1);
1146     }
1147     else {
1148         int r;
1149         r = gv(RC_INT); /* allocation size */
1150         /* sub r,%rsp */
1151         o(0x2b);
1152         o(0xe0 | r);
1153         /* We align to 16 bytes rather than align */
1154         /* and ~15, %esp */
1155         o(0xf0e483);
1156         vpop();
1157     }
1158 }
1159
1160 /* end of X86 code generator */
1161 /*************************************************************/
1162 #endif
1163 /*************************************************************/