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_ST0     0x0008
  35 #define RC_ECX     0x0010
  36 #define RC_EDX     0x0020
  37 #define RC_EBX     0x0040
  38
  39 #define RC_IRET    RC_EAX /* function return: integer register */
  40 #define RC_LRET    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_LRET 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 /******************************************************/
  75 #else /* ! TARGET_DEFS_ONLY */
  76 /******************************************************/
  77 #include "tcc.h"
  78
  79 /* define to 1/0 to [not] have EBX as 4th register */
  80 #define USE_EBX 0
  81
  82 ST_DATA const int reg_classes[NB_REGS] = {
  83     /* eax */ RC_INT | RC_EAX,
  84     /* ecx */ RC_INT | RC_ECX,
  85     /* edx */ RC_INT | RC_EDX,
  86     /* ebx */ (RC_INT | RC_EBX) * USE_EBX,
  87     /* st0 */ RC_FLOAT | RC_ST0,
  88 };
  89
  90 static unsigned long func_sub_sp_offset;
  91 static int func_ret_sub;
  92 #ifdef CONFIG_TCC_BCHECK
  93 static addr_t func_bound_offset;
  94 static unsigned long func_bound_ind;
  95 #endif
  96
  97 /* XXX: make it faster ? */
  98 ST_FUNC void g(int c)
  99 {
 100     int ind1;
 101     if (nocode_wanted)
 102         return;
 103     ind1 = ind + 1;
 104     if (ind1 > cur_text_section->data_allocated)
 105         section_realloc(cur_text_section, ind1);
 106     cur_text_section->data[ind] = c;
 107     ind = ind1;
 108 }
 109
 110 ST_FUNC void o(unsigned int c)
 111 {
 112     while (c) {
 113         g(c);
 114         c = c >> 8;
 115     }
 116 }
 117
 118 ST_FUNC void gen_le16(int v)
 119 {
 120     g(v);
 121     g(v >> 8);
 122 }
 123
 124 ST_FUNC void gen_le32(int c)
 125 {
 126     g(c);
 127     g(c >> 8);
 128     g(c >> 16);
 129     g(c >> 24);
 130 }
 131
 132 /* output a symbol and patch all calls to it */
 133 ST_FUNC void gsym_addr(int t, int a)
 134 {
 135     while (t) {
 136         unsigned char *ptr = cur_text_section->data + t;
 137         uint32_t n = read32le(ptr); /* next value */
 138         write32le(ptr, a - t - 4);
 139         t = n;
 140     }
 141 }
 142
 143 ST_FUNC void gsym(int t)
 144 {
 145     gsym_addr(t, ind);
 146 }
 147
 148 /* instruction + 4 bytes data. Return the address of the data */
 149 static int oad(int c, int s)
 150 {
 151     int t;
 152     if (nocode_wanted)
 153         return s;
 154     o(c);
 155     t = ind;
 156     gen_le32(s);
 157     return t;
 158 }
 159
 160 ST_FUNC void gen_fill_nops(int bytes)
 161 {
 162     while (bytes--)
 163       g(0x90);
 164 }
 165
 166 /* generate jmp to a label */
 167 #define gjmp2(instr,lbl) oad(instr,lbl)
 168
 169 /* output constant with relocation if 'r & VT_SYM' is true */
 170 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 171 {
 172     if (r & VT_SYM)
 173         greloc(cur_text_section, sym, ind, R_386_32);
 174     gen_le32(c);
 175 }
 176
 177 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 178 {
 179     if (r & VT_SYM)
 180         greloc(cur_text_section, sym, ind, R_386_PC32);
 181     gen_le32(c - 4);
 182 }
 183
 184 /* generate a modrm reference. 'op_reg' contains the additional 3
 185    opcode bits */
 186 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 187 {
 188     op_reg = op_reg << 3;
 189     if ((r & VT_VALMASK) == VT_CONST) {
 190         /* constant memory reference */
 191         o(0x05 | op_reg);
 192         gen_addr32(r, sym, c);
 193     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 194         /* currently, we use only ebp as base */
 195         if (c == (char)c) {
 196             /* short reference */
 197             o(0x45 | op_reg);
 198             g(c);
 199         } else {
 200             oad(0x85 | op_reg, c);
 201         }
 202     } else {
 203         g(0x00 | op_reg | (r & VT_VALMASK));
 204     }
 205 }
 206
 207 /* load 'r' from value 'sv' */
 208 ST_FUNC void load(int r, SValue *sv)
 209 {
 210     int v, t, ft, fc, fr;
 211     SValue v1;
 212
 213 #ifdef TCC_TARGET_PE
 214     SValue v2;
 215     sv = pe_getimport(sv, &v2);
 216 #endif
 217
 218     fr = sv->r;
 219     ft = sv->type.t & ~VT_DEFSIGN;
 220     fc = sv->c.i;
 221
 222     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 223
 224     v = fr & VT_VALMASK;
 225     if (fr & VT_LVAL) {
 226         if (v == VT_LLOCAL) {
 227             v1.type.t = VT_INT;
 228             v1.r = VT_LOCAL | VT_LVAL;
 229             v1.c.i = fc;
 230             fr = r;
 231             if (!(reg_classes[fr] & RC_INT))
 232                 fr = get_reg(RC_INT);
 233             load(fr, &v1);
 234         }
 235         if ((ft & VT_BTYPE) == VT_FLOAT) {
 236             o(0xd9); /* flds */
 237             r = 0;
 238         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 239             o(0xdd); /* fldl */
 240             r = 0;
 241         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 242             o(0xdb); /* fldt */
 243             r = 5;
 244         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 245             o(0xbe0f);   /* movsbl */
 246         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 247             o(0xb60f);   /* movzbl */
 248         } else if ((ft & VT_TYPE) == VT_SHORT) {
 249             o(0xbf0f);   /* movswl */
 250         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 251             o(0xb70f);   /* movzwl */
 252         } else {
 253             o(0x8b);     /* movl */
 254         }
 255         gen_modrm(r, fr, sv->sym, fc);
 256     } else {
 257         if (v == VT_CONST) {
 258             o(0xb8 + r); /* mov $xx, r */
 259             gen_addr32(fr, sv->sym, fc);
 260         } else if (v == VT_LOCAL) {
 261             if (fc) {
 262                 o(0x8d); /* lea xxx(%ebp), r */
 263                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 264             } else {
 265                 o(0x89);
 266                 o(0xe8 + r); /* mov %ebp, r */
 267             }
 268         } else if (v == VT_CMP) {
 269             oad(0xb8 + r, 0); /* mov $0, r */
 270             o(0x0f); /* setxx %br */
 271             o(fc);
 272             o(0xc0 + r);
 273         } else if (v == VT_JMP || v == VT_JMPI) {
 274             t = v & 1;
 275             oad(0xb8 + r, t); /* mov $1, r */
 276             o(0x05eb); /* jmp after */
 277             gsym(fc);
 278             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 279         } else if (v != r) {
 280             o(0x89);
 281             o(0xc0 + r + v * 8); /* mov v, r */
 282         }
 283     }
 284 }
 285
 286 /* store register 'r' in lvalue 'v' */
 287 ST_FUNC void store(int r, SValue *v)
 288 {
 289     int fr, bt, ft, fc;
 290
 291 #ifdef TCC_TARGET_PE
 292     SValue v2;
 293     v = pe_getimport(v, &v2);
 294 #endif
 295
 296     ft = v->type.t;
 297     fc = v->c.i;
 298     fr = v->r & VT_VALMASK;
 299     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 300     bt = ft & VT_BTYPE;
 301     /* XXX: incorrect if float reg to reg */
 302     if (bt == VT_FLOAT) {
 303         o(0xd9); /* fsts */
 304         r = 2;
 305     } else if (bt == VT_DOUBLE) {
 306         o(0xdd); /* fstpl */
 307         r = 2;
 308     } else if (bt == VT_LDOUBLE) {
 309         o(0xc0d9); /* fld %st(0) */
 310         o(0xdb); /* fstpt */
 311         r = 7;
 312     } else {
 313         if (bt == VT_SHORT)
 314             o(0x66);
 315         if (bt == VT_BYTE || bt == VT_BOOL)
 316             o(0x88);
 317         else
 318             o(0x89);
 319     }
 320     if (fr == VT_CONST ||
 321         fr == VT_LOCAL ||
 322         (v->r & VT_LVAL)) {
 323         gen_modrm(r, v->r, v->sym, fc);
 324     } else if (fr != r) {
 325         o(0xc0 + fr + r * 8); /* mov r, fr */
 326     }
 327 }
 328
 329 static void gadd_sp(int val)
 330 {
 331     if (val == (char)val) {
 332         o(0xc483);
 333         g(val);
 334     } else {
 335         oad(0xc481, val); /* add $xxx, %esp */
 336     }
 337 }
 338
 339 #if defined CONFIG_TCC_BCHECK || defined TCC_TARGET_PE
 340 static void gen_static_call(int v)
 341 {
 342     Sym *sym;
 343
 344     sym = external_global_sym(v, &func_old_type);
 345     oad(0xe8, -4);
 346     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 347 }
 348 #endif
 349
 350 /* 'is_jmp' is '1' if it is a jump */
 351 static void gcall_or_jmp(int is_jmp)
 352 {
 353     int r;
 354     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST && (vtop->r & VT_SYM)) {
 355         /* constant and relocation case */
 356         greloc(cur_text_section, vtop->sym, ind + 1, R_386_PC32);
 357         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 358     } else {
 359         /* otherwise, indirect call */
 360         r = gv(RC_INT);
 361         o(0xff); /* call/jmp *r */
 362         o(0xd0 + r + (is_jmp << 4));
 363     }
 364     if (!is_jmp) {
 365         int rt;
 366         /* extend the return value to the whole register if necessary
 367            visual studio and gcc do not always set the whole eax register
 368            when assigning the return value of a function  */
 369         rt = vtop->type.ref->type.t;
 370         switch (rt & VT_BTYPE) {
 371             case VT_BYTE:
 372             case VT_BOOL:
 373                 if (rt & VT_UNSIGNED) {
 374                     o(0xc0b60f); /* movzx %al, %eax */
 375                 }
 376                 else {
 377                     o(0xc0be0f); /* movsx %al, %eax */
 378                 }
 379                 break;
 380             case VT_SHORT:
 381                 if (rt & VT_UNSIGNED) {
 382                     o(0xc0b70f); /* movzx %ax, %eax */
 383                 }
 384                 else {
 385                     o(0xc0bf0f); /* movsx %ax, %eax */
 386                 }
 387                 break;
 388             default:
 389                 break;
 390         }
 391     }
 392 }
 393
 394 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 395 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 396
 397 /* Return the number of registers needed to return the struct, or 0 if
 398    returning via struct pointer. */
 399 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 400 {
 401 #ifdef TCC_TARGET_PE
 402     int size, align;
 403     *ret_align = 1; // Never have to re-align return values for x86
 404     *regsize = 4;
 405     size = type_size(vt, &align);
 406     if (size > 8 || (size & (size - 1)))
 407         return 0;
 408     if (size == 8)
 409         ret->t = VT_LLONG;
 410     else if (size == 4)
 411         ret->t = VT_INT;
 412     else if (size == 2)
 413         ret->t = VT_SHORT;
 414     else
 415         ret->t = VT_BYTE;
 416     ret->ref = NULL;
 417     return 1;
 418 #else
 419     *ret_align = 1; // Never have to re-align return values for x86
 420     return 0;
 421 #endif
 422 }
 423
 424 /* Generate function call. The function address is pushed first, then
 425    all the parameters in call order. This functions pops all the
 426    parameters and the function address. */
 427 ST_FUNC void gfunc_call(int nb_args)
 428 {
 429     int size, align, r, args_size, i, func_call;
 430     Sym *func_sym;
 431
 432     args_size = 0;
 433     for(i = 0;i < nb_args; i++) {
 434         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 435             size = type_size(&vtop->type, &align);
 436             /* align to stack align size */
 437             size = (size + 3) & ~3;
 438             /* allocate the necessary size on stack */
 439             oad(0xec81, size); /* sub $xxx, %esp */
 440             /* generate structure store */
 441             r = get_reg(RC_INT);
 442             o(0x89); /* mov %esp, r */
 443             o(0xe0 + r);
 444             vset(&vtop->type, r | VT_LVAL, 0);
 445             vswap();
 446             vstore();
 447             args_size += size;
 448         } else if (is_float(vtop->type.t)) {
 449             gv(RC_FLOAT); /* only one float register */
 450             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 451                 size = 4;
 452             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 453                 size = 8;
 454             else
 455                 size = 12;
 456             oad(0xec81, size); /* sub $xxx, %esp */
 457             if (size == 12)
 458                 o(0x7cdb);
 459             else
 460                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 461             g(0x24);
 462             g(0x00);
 463             args_size += size;
 464         } else {
 465             /* simple type (currently always same size) */
 466             /* XXX: implicit cast ? */
 467             r = gv(RC_INT);
 468             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 469                 size = 8;
 470                 o(0x50 + vtop->r2); /* push r */
 471             } else {
 472                 size = 4;
 473             }
 474             o(0x50 + r); /* push r */
 475             args_size += size;
 476         }
 477         vtop--;
 478     }
 479     save_regs(0); /* save used temporary registers */
 480     func_sym = vtop->type.ref;
 481     func_call = func_sym->f.func_call;
 482     /* fast call case */
 483     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 484         func_call == FUNC_FASTCALLW) {
 485         int fastcall_nb_regs;
 486         uint8_t *fastcall_regs_ptr;
 487         if (func_call == FUNC_FASTCALLW) {
 488             fastcall_regs_ptr = fastcallw_regs;
 489             fastcall_nb_regs = 2;
 490         } else {
 491             fastcall_regs_ptr = fastcall_regs;
 492             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 493         }
 494         for(i = 0;i < fastcall_nb_regs; i++) {
 495             if (args_size <= 0)
 496                 break;
 497             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 498             /* XXX: incorrect for struct/floats */
 499             args_size -= 4;
 500         }
 501     }
 502 #ifndef TCC_TARGET_PE
 503     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 504         args_size -= 4;
 505 #endif
 506     gcall_or_jmp(0);
 507
 508     if (args_size && func_call != FUNC_STDCALL && func_call != FUNC_FASTCALLW)
 509         gadd_sp(args_size);
 510     vtop--;
 511 }
 512
 513 #ifdef TCC_TARGET_PE
 514 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 515 #else
 516 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 517 #endif
 518
 519 /* generate function prolog of type 't' */
 520 ST_FUNC void gfunc_prolog(CType *func_type)
 521 {
 522     int addr, align, size, func_call, fastcall_nb_regs;
 523     int param_index, param_addr;
 524     uint8_t *fastcall_regs_ptr;
 525     Sym *sym;
 526     CType *type;
 527
 528     sym = func_type->ref;
 529     func_call = sym->f.func_call;
 530     addr = 8;
 531     loc = 0;
 532     func_vc = 0;
 533
 534     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 535         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 536         fastcall_regs_ptr = fastcall_regs;
 537     } else if (func_call == FUNC_FASTCALLW) {
 538         fastcall_nb_regs = 2;
 539         fastcall_regs_ptr = fastcallw_regs;
 540     } else {
 541         fastcall_nb_regs = 0;
 542         fastcall_regs_ptr = NULL;
 543     }
 544     param_index = 0;
 545
 546     ind += FUNC_PROLOG_SIZE;
 547     func_sub_sp_offset = ind;
 548     /* if the function returns a structure, then add an
 549        implicit pointer parameter */
 550     func_vt = sym->type;
 551     func_var = (sym->f.func_type == FUNC_ELLIPSIS);
 552 #ifdef TCC_TARGET_PE
 553     size = type_size(&func_vt,&align);
 554     if (((func_vt.t & VT_BTYPE) == VT_STRUCT)
 555         && (size > 8 || (size & (size - 1)))) {
 556 #else
 557     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 558 #endif
 559         /* XXX: fastcall case ? */
 560         func_vc = addr;
 561         addr += 4;
 562         param_index++;
 563     }
 564     /* define parameters */
 565     while ((sym = sym->next) != NULL) {
 566         type = &sym->type;
 567         size = type_size(type, &align);
 568         size = (size + 3) & ~3;
 569 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 570         /* structs are passed as pointer */
 571         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 572             size = 4;
 573         }
 574 #endif
 575         if (param_index < fastcall_nb_regs) {
 576             /* save FASTCALL register */
 577             loc -= 4;
 578             o(0x89);     /* movl */
 579             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 580             param_addr = loc;
 581         } else {
 582             param_addr = addr;
 583             addr += size;
 584         }
 585         sym_push(sym->v & ~SYM_FIELD, type,
 586                  VT_LOCAL | lvalue_type(type->t), param_addr);
 587         param_index++;
 588     }
 589     func_ret_sub = 0;
 590     /* pascal type call or fastcall ? */
 591     if (func_call == FUNC_STDCALL || func_call == FUNC_FASTCALLW)
 592         func_ret_sub = addr - 8;
 593 #ifndef TCC_TARGET_PE
 594     else if (func_vc)
 595         func_ret_sub = 4;
 596 #endif
 597
 598 #ifdef CONFIG_TCC_BCHECK
 599     /* leave some room for bound checking code */
 600     if (tcc_state->do_bounds_check) {
 601         func_bound_offset = lbounds_section->data_offset;
 602         func_bound_ind = ind;
 603         oad(0xb8, 0); /* lbound section pointer */
 604         oad(0xb8, 0); /* call to function */
 605     }
 606 #endif
 607 }
 608
 609 /* generate function epilog */
 610 ST_FUNC void gfunc_epilog(void)
 611 {
 612     addr_t v, saved_ind;
 613
 614 #ifdef CONFIG_TCC_BCHECK
 615     if (tcc_state->do_bounds_check
 616      && func_bound_offset != lbounds_section->data_offset) {
 617         addr_t saved_ind;
 618         addr_t *bounds_ptr;
 619         Sym *sym_data;
 620
 621         /* add end of table info */
 622         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 623         *bounds_ptr = 0;
 624
 625         /* generate bound local allocation */
 626         saved_ind = ind;
 627         ind = func_bound_ind;
 628         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 629                                func_bound_offset, lbounds_section->data_offset);
 630         greloc(cur_text_section, sym_data,
 631                ind + 1, R_386_32);
 632         oad(0xb8, 0); /* mov %eax, xxx */
 633         gen_static_call(TOK___bound_local_new);
 634         ind = saved_ind;
 635
 636         /* generate bound check local freeing */
 637         o(0x5250); /* save returned value, if any */
 638         greloc(cur_text_section, sym_data, ind + 1, R_386_32);
 639         oad(0xb8, 0); /* mov %eax, xxx */
 640         gen_static_call(TOK___bound_local_delete);
 641         o(0x585a); /* restore returned value, if any */
 642     }
 643 #endif
 644
 645     /* align local size to word & save local variables */
 646     v = (-loc + 3) & -4;
 647
 648 #if USE_EBX
 649     o(0x8b);
 650     gen_modrm(TREG_EBX, VT_LOCAL, NULL, -(v+4));
 651 #endif
 652
 653     o(0xc9); /* leave */
 654     if (func_ret_sub == 0) {
 655         o(0xc3); /* ret */
 656     } else {
 657         o(0xc2); /* ret n */
 658         g(func_ret_sub);
 659         g(func_ret_sub >> 8);
 660     }
 661     saved_ind = ind;
 662     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 663 #ifdef TCC_TARGET_PE
 664     if (v >= 4096) {
 665         oad(0xb8, v); /* mov stacksize, %eax */
 666         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 667     } else
 668 #endif
 669     {
 670         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 671         o(0xec81);  /* sub esp, stacksize */
 672         gen_le32(v);
 673 #ifdef TCC_TARGET_PE
 674         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 675 #endif
 676     }
 677     o(0x53 * USE_EBX); /* push ebx */
 678     ind = saved_ind;
 679 }
 680
 681 /* generate a jump to a label */
 682 ST_FUNC int gjmp(int t)
 683 {
 684     return gjmp2(0xe9, t);
 685 }
 686
 687 /* generate a jump to a fixed address */
 688 ST_FUNC void gjmp_addr(int a)
 689 {
 690     int r;
 691     r = a - ind - 2;
 692     if (r == (char)r) {
 693         g(0xeb);
 694         g(r);
 695     } else {
 696         oad(0xe9, a - ind - 5);
 697     }
 698 }
 699
 700 ST_FUNC void gtst_addr(int inv, int a)
 701 {
 702     int v = vtop->r & VT_VALMASK;
 703     if (v == VT_CMP) {
 704         inv ^= (vtop--)->c.i;
 705         a -= ind + 2;
 706         if (a == (char)a) {
 707             g(inv - 32);
 708             g(a);
 709         } else {
 710             g(0x0f);
 711             oad(inv - 16, a - 4);
 712         }
 713     } else if ((v & ~1) == VT_JMP) {
 714         if ((v & 1) != inv) {
 715             gjmp_addr(a);
 716             gsym(vtop->c.i);
 717         } else {
 718             gsym(vtop->c.i);
 719             o(0x05eb);
 720             gjmp_addr(a);
 721         }
 722         vtop--;
 723     }
 724 }
 725
 726 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 727 ST_FUNC int gtst(int inv, int t)
 728 {
 729     int v = vtop->r & VT_VALMASK;
 730     if (nocode_wanted) {
 731         ;
 732     } else if (v == VT_CMP) {
 733         /* fast case : can jump directly since flags are set */
 734         g(0x0f);
 735         t = gjmp2((vtop->c.i - 16) ^ inv, t);
 736     } else if (v == VT_JMP || v == VT_JMPI) {
 737         /* && or || optimization */
 738         if ((v & 1) == inv) {
 739             /* insert vtop->c jump list in t */
 740             uint32_t n1, n = vtop->c.i;
 741             if (n) {
 742                 while ((n1 = read32le(cur_text_section->data + n)))
 743                     n = n1;
 744                 write32le(cur_text_section->data + n, t);
 745                 t = vtop->c.i;
 746             }
 747         } else {
 748             t = gjmp(t);
 749             gsym(vtop->c.i);
 750         }
 751     }
 752     vtop--;
 753     return t;
 754 }
 755
 756 /* generate an integer binary operation */
 757 ST_FUNC void gen_opi(int op)
 758 {
 759     int r, fr, opc, c;
 760
 761     switch(op) {
 762     case '+':
 763     case TOK_ADDC1: /* add with carry generation */
 764         opc = 0;
 765     gen_op8:
 766         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 767             /* constant case */
 768             vswap();
 769             r = gv(RC_INT);
 770             vswap();
 771             c = vtop->c.i;
 772             if (c == (char)c) {
 773                 /* generate inc and dec for smaller code */
 774                 if (c==1 && opc==0 && op != TOK_ADDC1) {
 775                     o (0x40 | r); // inc
 776                 } else if (c==1 && opc==5 && op != TOK_SUBC1) {
 777                     o (0x48 | r); // dec
 778                 } else {
 779                     o(0x83);
 780                     o(0xc0 | (opc << 3) | r);
 781                     g(c);
 782                 }
 783             } else {
 784                 o(0x81);
 785                 oad(0xc0 | (opc << 3) | r, c);
 786             }
 787         } else {
 788             gv2(RC_INT, RC_INT);
 789             r = vtop[-1].r;
 790             fr = vtop[0].r;
 791             o((opc << 3) | 0x01);
 792             o(0xc0 + r + fr * 8);
 793         }
 794         vtop--;
 795         if (op >= TOK_ULT && op <= TOK_GT) {
 796             vtop->r = VT_CMP;
 797             vtop->c.i = op;
 798         }
 799         break;
 800     case '-':
 801     case TOK_SUBC1: /* sub with carry generation */
 802         opc = 5;
 803         goto gen_op8;
 804     case TOK_ADDC2: /* add with carry use */
 805         opc = 2;
 806         goto gen_op8;
 807     case TOK_SUBC2: /* sub with carry use */
 808         opc = 3;
 809         goto gen_op8;
 810     case '&':
 811         opc = 4;
 812         goto gen_op8;
 813     case '^':
 814         opc = 6;
 815         goto gen_op8;
 816     case '|':
 817         opc = 1;
 818         goto gen_op8;
 819     case '*':
 820         gv2(RC_INT, RC_INT);
 821         r = vtop[-1].r;
 822         fr = vtop[0].r;
 823         vtop--;
 824         o(0xaf0f); /* imul fr, r */
 825         o(0xc0 + fr + r * 8);
 826         break;
 827     case TOK_SHL:
 828         opc = 4;
 829         goto gen_shift;
 830     case TOK_SHR:
 831         opc = 5;
 832         goto gen_shift;
 833     case TOK_SAR:
 834         opc = 7;
 835     gen_shift:
 836         opc = 0xc0 | (opc << 3);
 837         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 838             /* constant case */
 839             vswap();
 840             r = gv(RC_INT);
 841             vswap();
 842             c = vtop->c.i & 0x1f;
 843             o(0xc1); /* shl/shr/sar $xxx, r */
 844             o(opc | r);
 845             g(c);
 846         } else {
 847             /* we generate the shift in ecx */
 848             gv2(RC_INT, RC_ECX);
 849             r = vtop[-1].r;
 850             o(0xd3); /* shl/shr/sar %cl, r */
 851             o(opc | r);
 852         }
 853         vtop--;
 854         break;
 855     case '/':
 856     case TOK_UDIV:
 857     case TOK_PDIV:
 858     case '%':
 859     case TOK_UMOD:
 860     case TOK_UMULL:
 861         /* first operand must be in eax */
 862         /* XXX: need better constraint for second operand */
 863         gv2(RC_EAX, RC_ECX);
 864         r = vtop[-1].r;
 865         fr = vtop[0].r;
 866         vtop--;
 867         save_reg(TREG_EDX);
 868         /* save EAX too if used otherwise */
 869         save_reg_upstack(TREG_EAX, 1);
 870         if (op == TOK_UMULL) {
 871             o(0xf7); /* mul fr */
 872             o(0xe0 + fr);
 873             vtop->r2 = TREG_EDX;
 874             r = TREG_EAX;
 875         } else {
 876             if (op == TOK_UDIV || op == TOK_UMOD) {
 877                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 878                 o(0xf0 + fr);
 879             } else {
 880                 o(0xf799); /* cltd, idiv fr, %eax */
 881                 o(0xf8 + fr);
 882             }
 883             if (op == '%' || op == TOK_UMOD)
 884                 r = TREG_EDX;
 885             else
 886                 r = TREG_EAX;
 887         }
 888         vtop->r = r;
 889         break;
 890     default:
 891         opc = 7;
 892         goto gen_op8;
 893     }
 894 }
 895
 896 /* generate a floating point operation 'v = t1 op t2' instruction. The
 897    two operands are guaranteed to have the same floating point type */
 898 /* XXX: need to use ST1 too */
 899 ST_FUNC void gen_opf(int op)
 900 {
 901     int a, ft, fc, swapped, r;
 902
 903     /* convert constants to memory references */
 904     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 905         vswap();
 906         gv(RC_FLOAT);
 907         vswap();
 908     }
 909     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 910         gv(RC_FLOAT);
 911
 912     /* must put at least one value in the floating point register */
 913     if ((vtop[-1].r & VT_LVAL) &&
 914         (vtop[0].r & VT_LVAL)) {
 915         vswap();
 916         gv(RC_FLOAT);
 917         vswap();
 918     }
 919     swapped = 0;
 920     /* swap the stack if needed so that t1 is the register and t2 is
 921        the memory reference */
 922     if (vtop[-1].r & VT_LVAL) {
 923         vswap();
 924         swapped = 1;
 925     }
 926     if (op >= TOK_ULT && op <= TOK_GT) {
 927         /* load on stack second operand */
 928         load(TREG_ST0, vtop);
 929         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 930         if (op == TOK_GE || op == TOK_GT)
 931             swapped = !swapped;
 932         else if (op == TOK_EQ || op == TOK_NE)
 933             swapped = 0;
 934         if (swapped)
 935             o(0xc9d9); /* fxch %st(1) */
 936         if (op == TOK_EQ || op == TOK_NE)
 937             o(0xe9da); /* fucompp */
 938         else
 939             o(0xd9de); /* fcompp */
 940         o(0xe0df); /* fnstsw %ax */
 941         if (op == TOK_EQ) {
 942             o(0x45e480); /* and $0x45, %ah */
 943             o(0x40fC80); /* cmp $0x40, %ah */
 944         } else if (op == TOK_NE) {
 945             o(0x45e480); /* and $0x45, %ah */
 946             o(0x40f480); /* xor $0x40, %ah */
 947             op = TOK_NE;
 948         } else if (op == TOK_GE || op == TOK_LE) {
 949             o(0x05c4f6); /* test $0x05, %ah */
 950             op = TOK_EQ;
 951         } else {
 952             o(0x45c4f6); /* test $0x45, %ah */
 953             op = TOK_EQ;
 954         }
 955         vtop--;
 956         vtop->r = VT_CMP;
 957         vtop->c.i = op;
 958     } else {
 959         /* no memory reference possible for long double operations */
 960         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 961             load(TREG_ST0, vtop);
 962             swapped = !swapped;
 963         }
 964
 965         switch(op) {
 966         default:
 967         case '+':
 968             a = 0;
 969             break;
 970         case '-':
 971             a = 4;
 972             if (swapped)
 973                 a++;
 974             break;
 975         case '*':
 976             a = 1;
 977             break;
 978         case '/':
 979             a = 6;
 980             if (swapped)
 981                 a++;
 982             break;
 983         }
 984         ft = vtop->type.t;
 985         fc = vtop->c.i;
 986         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 987             o(0xde); /* fxxxp %st, %st(1) */
 988             o(0xc1 + (a << 3));
 989         } else {
 990             /* if saved lvalue, then we must reload it */
 991             r = vtop->r;
 992             if ((r & VT_VALMASK) == VT_LLOCAL) {
 993                 SValue v1;
 994                 r = get_reg(RC_INT);
 995                 v1.type.t = VT_INT;
 996                 v1.r = VT_LOCAL | VT_LVAL;
 997                 v1.c.i = fc;
 998                 load(r, &v1);
 999                 fc = 0;
1000             }
1001
1002             if ((ft & VT_BTYPE) == VT_DOUBLE)
1003                 o(0xdc);
1004             else
1005                 o(0xd8);
1006             gen_modrm(a, r, vtop->sym, fc);
1007         }
1008         vtop--;
1009     }
1010 }
1011
1012 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1013    and 'long long' cases. */
1014 ST_FUNC void gen_cvt_itof(int t)
1015 {
1016     save_reg(TREG_ST0);
1017     gv(RC_INT);
1018     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
1019         /* signed long long to float/double/long double (unsigned case
1020            is handled generically) */
1021         o(0x50 + vtop->r2); /* push r2 */
1022         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1023         o(0x242cdf); /* fildll (%esp) */
1024         o(0x08c483); /* add $8, %esp */
1025     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
1026                (VT_INT | VT_UNSIGNED)) {
1027         /* unsigned int to float/double/long double */
1028         o(0x6a); /* push $0 */
1029         g(0x00);
1030         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1031         o(0x242cdf); /* fildll (%esp) */
1032         o(0x08c483); /* add $8, %esp */
1033     } else {
1034         /* int to float/double/long double */
1035         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1036         o(0x2404db); /* fildl (%esp) */
1037         o(0x04c483); /* add $4, %esp */
1038     }
1039     vtop->r = TREG_ST0;
1040 }
1041
1042 /* convert fp to int 't' type */
1043 ST_FUNC void gen_cvt_ftoi(int t)
1044 {
1045     int bt = vtop->type.t & VT_BTYPE;
1046     if (bt == VT_FLOAT)
1047         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1048     else if (bt == VT_LDOUBLE)
1049         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1050     else
1051         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1052     vswap();
1053     gfunc_call(1);
1054     vpushi(0);
1055     vtop->r = REG_IRET;
1056     vtop->r2 = REG_LRET;
1057 }
1058
1059 /* convert from one floating point type to another */
1060 ST_FUNC void gen_cvt_ftof(int t)
1061 {
1062     /* all we have to do on i386 is to put the float in a register */
1063     gv(RC_FLOAT);
1064 }
1065
1066 /* computed goto support */
1067 ST_FUNC void ggoto(void)
1068 {
1069     gcall_or_jmp(1);
1070     vtop--;
1071 }
1072
1073 /* bound check support functions */
1074 #ifdef CONFIG_TCC_BCHECK
1075
1076 /* generate a bounded pointer addition */
1077 ST_FUNC void gen_bounded_ptr_add(void)
1078 {
1079     /* prepare fast i386 function call (args in eax and edx) */
1080     gv2(RC_EAX, RC_EDX);
1081     /* save all temporary registers */
1082     vtop -= 2;
1083     save_regs(0);
1084     /* do a fast function call */
1085     gen_static_call(TOK___bound_ptr_add);
1086     /* returned pointer is in eax */
1087     vtop++;
1088     vtop->r = TREG_EAX | VT_BOUNDED;
1089     /* address of bounding function call point */
1090     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1091 }
1092
1093 /* patch pointer addition in vtop so that pointer dereferencing is
1094    also tested */
1095 ST_FUNC void gen_bounded_ptr_deref(void)
1096 {
1097     addr_t func;
1098     int  size, align;
1099     Elf32_Rel *rel;
1100     Sym *sym;
1101
1102     size = 0;
1103     /* XXX: put that code in generic part of tcc */
1104     if (!is_float(vtop->type.t)) {
1105         if (vtop->r & VT_LVAL_BYTE)
1106             size = 1;
1107         else if (vtop->r & VT_LVAL_SHORT)
1108             size = 2;
1109     }
1110     if (!size)
1111         size = type_size(&vtop->type, &align);
1112     switch(size) {
1113     case  1: func = TOK___bound_ptr_indir1; break;
1114     case  2: func = TOK___bound_ptr_indir2; break;
1115     case  4: func = TOK___bound_ptr_indir4; break;
1116     case  8: func = TOK___bound_ptr_indir8; break;
1117     case 12: func = TOK___bound_ptr_indir12; break;
1118     case 16: func = TOK___bound_ptr_indir16; break;
1119     default:
1120         tcc_error("unhandled size when dereferencing bounded pointer");
1121         func = 0;
1122         break;
1123     }
1124
1125     /* patch relocation */
1126     /* XXX: find a better solution ? */
1127     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1128     sym = external_global_sym(func, &func_old_type);
1129     if (!sym->c)
1130         put_extern_sym(sym, NULL, 0, 0);
1131     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1132 }
1133 #endif
1134
1135 /* Save the stack pointer onto the stack */
1136 ST_FUNC void gen_vla_sp_save(int addr) {
1137     /* mov %esp,addr(%ebp)*/
1138     o(0x89);
1139     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1140 }
1141
1142 /* Restore the SP from a location on the stack */
1143 ST_FUNC void gen_vla_sp_restore(int addr) {
1144     o(0x8b);
1145     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1146 }
1147
1148 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1149 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1150 #ifdef TCC_TARGET_PE
1151     /* alloca does more than just adjust %rsp on Windows */
1152     vpush_global_sym(&func_old_type, TOK_alloca);
1153     vswap(); /* Move alloca ref past allocation size */
1154     gfunc_call(1);
1155 #else
1156     int r;
1157     r = gv(RC_INT); /* allocation size */
1158     /* sub r,%rsp */
1159     o(0x2b);
1160     o(0xe0 | r);
1161     /* We align to 16 bytes rather than align */
1162     /* and ~15, %esp */
1163     o(0xf0e483);
1164     vpop();
1165 #endif
1166 }
1167
1168 /* end of X86 code generator */
1169 /*************************************************************/
1170 #endif
1171 /*************************************************************/