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, 0);
 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                 if (rt & VT_UNSIGNED) {
 373                     o(0xc0b60f); /* movzx %al, %eax */
 374                 }
 375                 else {
 376                     o(0xc0be0f); /* movsx %al, %eax */
 377                 }
 378                 break;
 379             case VT_SHORT:
 380                 if (rt & VT_UNSIGNED) {
 381                     o(0xc0b70f); /* movzx %ax, %eax */
 382                 }
 383                 else {
 384                     o(0xc0bf0f); /* movsx %ax, %eax */
 385                 }
 386                 break;
 387             default:
 388                 break;
 389         }
 390     }
 391 }
 392
 393 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 394 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 395
 396 /* Return the number of registers needed to return the struct, or 0 if
 397    returning via struct pointer. */
 398 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 399 {
 400 #ifdef TCC_TARGET_PE
 401     int size, align;
 402     *ret_align = 1; // Never have to re-align return values for x86
 403     *regsize = 4;
 404     size = type_size(vt, &align);
 405     if (size > 8 || (size & (size - 1)))
 406         return 0;
 407     if (size == 8)
 408         ret->t = VT_LLONG;
 409     else if (size == 4)
 410         ret->t = VT_INT;
 411     else if (size == 2)
 412         ret->t = VT_SHORT;
 413     else
 414         ret->t = VT_BYTE;
 415     ret->ref = NULL;
 416     return 1;
 417 #else
 418     *ret_align = 1; // Never have to re-align return values for x86
 419     return 0;
 420 #endif
 421 }
 422
 423 /* Generate function call. The function address is pushed first, then
 424    all the parameters in call order. This functions pops all the
 425    parameters and the function address. */
 426 ST_FUNC void gfunc_call(int nb_args)
 427 {
 428     int size, align, r, args_size, i, func_call;
 429     Sym *func_sym;
 430
 431     args_size = 0;
 432     for(i = 0;i < nb_args; i++) {
 433         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 434             size = type_size(&vtop->type, &align);
 435             /* align to stack align size */
 436             size = (size + 3) & ~3;
 437             /* allocate the necessary size on stack */
 438             oad(0xec81, size); /* sub $xxx, %esp */
 439             /* generate structure store */
 440             r = get_reg(RC_INT);
 441             o(0x89); /* mov %esp, r */
 442             o(0xe0 + r);
 443             vset(&vtop->type, r | VT_LVAL, 0);
 444             vswap();
 445             vstore();
 446             args_size += size;
 447         } else if (is_float(vtop->type.t)) {
 448             gv(RC_FLOAT); /* only one float register */
 449             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 450                 size = 4;
 451             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 452                 size = 8;
 453             else
 454                 size = 12;
 455             oad(0xec81, size); /* sub $xxx, %esp */
 456             if (size == 12)
 457                 o(0x7cdb);
 458             else
 459                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 460             g(0x24);
 461             g(0x00);
 462             args_size += size;
 463         } else {
 464             /* simple type (currently always same size) */
 465             /* XXX: implicit cast ? */
 466             r = gv(RC_INT);
 467             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 468                 size = 8;
 469                 o(0x50 + vtop->r2); /* push r */
 470             } else {
 471                 size = 4;
 472             }
 473             o(0x50 + r); /* push r */
 474             args_size += size;
 475         }
 476         vtop--;
 477     }
 478     save_regs(0); /* save used temporary registers */
 479     func_sym = vtop->type.ref;
 480     func_call = func_sym->f.func_call;
 481     /* fast call case */
 482     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 483         func_call == FUNC_FASTCALLW) {
 484         int fastcall_nb_regs;
 485         uint8_t *fastcall_regs_ptr;
 486         if (func_call == FUNC_FASTCALLW) {
 487             fastcall_regs_ptr = fastcallw_regs;
 488             fastcall_nb_regs = 2;
 489         } else {
 490             fastcall_regs_ptr = fastcall_regs;
 491             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 492         }
 493         for(i = 0;i < fastcall_nb_regs; i++) {
 494             if (args_size <= 0)
 495                 break;
 496             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 497             /* XXX: incorrect for struct/floats */
 498             args_size -= 4;
 499         }
 500     }
 501 #ifndef TCC_TARGET_PE
 502     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 503         args_size -= 4;
 504 #endif
 505     gcall_or_jmp(0);
 506
 507     if (args_size && func_call != FUNC_STDCALL && func_call != FUNC_FASTCALLW)
 508         gadd_sp(args_size);
 509     vtop--;
 510 }
 511
 512 #ifdef TCC_TARGET_PE
 513 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 514 #else
 515 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 516 #endif
 517
 518 /* generate function prolog of type 't' */
 519 ST_FUNC void gfunc_prolog(CType *func_type)
 520 {
 521     int addr, align, size, func_call, fastcall_nb_regs;
 522     int param_index, param_addr;
 523     uint8_t *fastcall_regs_ptr;
 524     Sym *sym;
 525     CType *type;
 526
 527     sym = func_type->ref;
 528     func_call = sym->f.func_call;
 529     addr = 8;
 530     loc = 0;
 531     func_vc = 0;
 532
 533     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 534         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 535         fastcall_regs_ptr = fastcall_regs;
 536     } else if (func_call == FUNC_FASTCALLW) {
 537         fastcall_nb_regs = 2;
 538         fastcall_regs_ptr = fastcallw_regs;
 539     } else {
 540         fastcall_nb_regs = 0;
 541         fastcall_regs_ptr = NULL;
 542     }
 543     param_index = 0;
 544
 545     ind += FUNC_PROLOG_SIZE;
 546     func_sub_sp_offset = ind;
 547     /* if the function returns a structure, then add an
 548        implicit pointer parameter */
 549     func_vt = sym->type;
 550     func_var = (sym->f.func_type == FUNC_ELLIPSIS);
 551 #ifdef TCC_TARGET_PE
 552     size = type_size(&func_vt,&align);
 553     if (((func_vt.t & VT_BTYPE) == VT_STRUCT)
 554         && (size > 8 || (size & (size - 1)))) {
 555 #else
 556     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 557 #endif
 558         /* XXX: fastcall case ? */
 559         func_vc = addr;
 560         addr += 4;
 561         param_index++;
 562     }
 563     /* define parameters */
 564     while ((sym = sym->next) != NULL) {
 565         type = &sym->type;
 566         size = type_size(type, &align);
 567         size = (size + 3) & ~3;
 568 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 569         /* structs are passed as pointer */
 570         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 571             size = 4;
 572         }
 573 #endif
 574         if (param_index < fastcall_nb_regs) {
 575             /* save FASTCALL register */
 576             loc -= 4;
 577             o(0x89);     /* movl */
 578             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 579             param_addr = loc;
 580         } else {
 581             param_addr = addr;
 582             addr += size;
 583         }
 584         sym_push(sym->v & ~SYM_FIELD, type,
 585                  VT_LOCAL | lvalue_type(type->t), param_addr);
 586         param_index++;
 587     }
 588     func_ret_sub = 0;
 589     /* pascal type call or fastcall ? */
 590     if (func_call == FUNC_STDCALL || func_call == FUNC_FASTCALLW)
 591         func_ret_sub = addr - 8;
 592 #ifndef TCC_TARGET_PE
 593     else if (func_vc)
 594         func_ret_sub = 4;
 595 #endif
 596
 597 #ifdef CONFIG_TCC_BCHECK
 598     /* leave some room for bound checking code */
 599     if (tcc_state->do_bounds_check) {
 600         func_bound_offset = lbounds_section->data_offset;
 601         func_bound_ind = ind;
 602         oad(0xb8, 0); /* lbound section pointer */
 603         oad(0xb8, 0); /* call to function */
 604     }
 605 #endif
 606 }
 607
 608 /* generate function epilog */
 609 ST_FUNC void gfunc_epilog(void)
 610 {
 611     addr_t v, saved_ind;
 612
 613 #ifdef CONFIG_TCC_BCHECK
 614     if (tcc_state->do_bounds_check
 615      && func_bound_offset != lbounds_section->data_offset) {
 616         addr_t saved_ind;
 617         addr_t *bounds_ptr;
 618         Sym *sym_data;
 619
 620         /* add end of table info */
 621         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 622         *bounds_ptr = 0;
 623
 624         /* generate bound local allocation */
 625         saved_ind = ind;
 626         ind = func_bound_ind;
 627         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 628                                func_bound_offset, lbounds_section->data_offset);
 629         greloc(cur_text_section, sym_data,
 630                ind + 1, R_386_32);
 631         oad(0xb8, 0); /* mov %eax, xxx */
 632         gen_static_call(TOK___bound_local_new);
 633         ind = saved_ind;
 634
 635         /* generate bound check local freeing */
 636         o(0x5250); /* save returned value, if any */
 637         greloc(cur_text_section, sym_data, ind + 1, R_386_32);
 638         oad(0xb8, 0); /* mov %eax, xxx */
 639         gen_static_call(TOK___bound_local_delete);
 640         o(0x585a); /* restore returned value, if any */
 641     }
 642 #endif
 643
 644     /* align local size to word & save local variables */
 645     v = (-loc + 3) & -4;
 646
 647 #if USE_EBX
 648     o(0x8b);
 649     gen_modrm(TREG_EBX, VT_LOCAL, NULL, -(v+4));
 650 #endif
 651
 652     o(0xc9); /* leave */
 653     if (func_ret_sub == 0) {
 654         o(0xc3); /* ret */
 655     } else {
 656         o(0xc2); /* ret n */
 657         g(func_ret_sub);
 658         g(func_ret_sub >> 8);
 659     }
 660     saved_ind = ind;
 661     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 662 #ifdef TCC_TARGET_PE
 663     if (v >= 4096) {
 664         oad(0xb8, v); /* mov stacksize, %eax */
 665         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 666     } else
 667 #endif
 668     {
 669         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 670         o(0xec81);  /* sub esp, stacksize */
 671         gen_le32(v);
 672 #ifdef TCC_TARGET_PE
 673         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 674 #endif
 675     }
 676     o(0x53 * USE_EBX); /* push ebx */
 677     ind = saved_ind;
 678 }
 679
 680 /* generate a jump to a label */
 681 ST_FUNC int gjmp(int t)
 682 {
 683     return gjmp2(0xe9, t);
 684 }
 685
 686 /* generate a jump to a fixed address */
 687 ST_FUNC void gjmp_addr(int a)
 688 {
 689     int r;
 690     r = a - ind - 2;
 691     if (r == (char)r) {
 692         g(0xeb);
 693         g(r);
 694     } else {
 695         oad(0xe9, a - ind - 5);
 696     }
 697 }
 698
 699 ST_FUNC void gtst_addr(int inv, int a)
 700 {
 701     int v = vtop->r & VT_VALMASK;
 702     if (v == VT_CMP) {
 703         inv ^= (vtop--)->c.i;
 704         a -= ind + 2;
 705         if (a == (char)a) {
 706             g(inv - 32);
 707             g(a);
 708         } else {
 709             g(0x0f);
 710             oad(inv - 16, a - 4);
 711         }
 712     } else if ((v & ~1) == VT_JMP) {
 713         if ((v & 1) != inv) {
 714             gjmp_addr(a);
 715             gsym(vtop->c.i);
 716         } else {
 717             gsym(vtop->c.i);
 718             o(0x05eb);
 719             gjmp_addr(a);
 720         }
 721         vtop--;
 722     }
 723 }
 724
 725 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 726 ST_FUNC int gtst(int inv, int t)
 727 {
 728     int v = vtop->r & VT_VALMASK;
 729     if (nocode_wanted) {
 730         ;
 731     } else if (v == VT_CMP) {
 732         /* fast case : can jump directly since flags are set */
 733         g(0x0f);
 734         t = gjmp2((vtop->c.i - 16) ^ inv, t);
 735     } else if (v == VT_JMP || v == VT_JMPI) {
 736         /* && or || optimization */
 737         if ((v & 1) == inv) {
 738             /* insert vtop->c jump list in t */
 739             uint32_t n1, n = vtop->c.i;
 740             if (n) {
 741                 while ((n1 = read32le(cur_text_section->data + n)))
 742                     n = n1;
 743                 write32le(cur_text_section->data + n, t);
 744                 t = vtop->c.i;
 745             }
 746         } else {
 747             t = gjmp(t);
 748             gsym(vtop->c.i);
 749         }
 750     }
 751     vtop--;
 752     return t;
 753 }
 754
 755 /* generate an integer binary operation */
 756 ST_FUNC void gen_opi(int op)
 757 {
 758     int r, fr, opc, c;
 759
 760     switch(op) {
 761     case '+':
 762     case TOK_ADDC1: /* add with carry generation */
 763         opc = 0;
 764     gen_op8:
 765         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 766             /* constant case */
 767             vswap();
 768             r = gv(RC_INT);
 769             vswap();
 770             c = vtop->c.i;
 771             if (c == (char)c) {
 772                 /* generate inc and dec for smaller code */
 773                 if (c==1 && opc==0 && op != TOK_ADDC1) {
 774                     o (0x40 | r); // inc
 775                 } else if (c==1 && opc==5 && op != TOK_SUBC1) {
 776                     o (0x48 | r); // dec
 777                 } else {
 778                     o(0x83);
 779                     o(0xc0 | (opc << 3) | r);
 780                     g(c);
 781                 }
 782             } else {
 783                 o(0x81);
 784                 oad(0xc0 | (opc << 3) | r, c);
 785             }
 786         } else {
 787             gv2(RC_INT, RC_INT);
 788             r = vtop[-1].r;
 789             fr = vtop[0].r;
 790             o((opc << 3) | 0x01);
 791             o(0xc0 + r + fr * 8);
 792         }
 793         vtop--;
 794         if (op >= TOK_ULT && op <= TOK_GT) {
 795             vtop->r = VT_CMP;
 796             vtop->c.i = op;
 797         }
 798         break;
 799     case '-':
 800     case TOK_SUBC1: /* sub with carry generation */
 801         opc = 5;
 802         goto gen_op8;
 803     case TOK_ADDC2: /* add with carry use */
 804         opc = 2;
 805         goto gen_op8;
 806     case TOK_SUBC2: /* sub with carry use */
 807         opc = 3;
 808         goto gen_op8;
 809     case '&':
 810         opc = 4;
 811         goto gen_op8;
 812     case '^':
 813         opc = 6;
 814         goto gen_op8;
 815     case '|':
 816         opc = 1;
 817         goto gen_op8;
 818     case '*':
 819         gv2(RC_INT, RC_INT);
 820         r = vtop[-1].r;
 821         fr = vtop[0].r;
 822         vtop--;
 823         o(0xaf0f); /* imul fr, r */
 824         o(0xc0 + fr + r * 8);
 825         break;
 826     case TOK_SHL:
 827         opc = 4;
 828         goto gen_shift;
 829     case TOK_SHR:
 830         opc = 5;
 831         goto gen_shift;
 832     case TOK_SAR:
 833         opc = 7;
 834     gen_shift:
 835         opc = 0xc0 | (opc << 3);
 836         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 837             /* constant case */
 838             vswap();
 839             r = gv(RC_INT);
 840             vswap();
 841             c = vtop->c.i & 0x1f;
 842             o(0xc1); /* shl/shr/sar $xxx, r */
 843             o(opc | r);
 844             g(c);
 845         } else {
 846             /* we generate the shift in ecx */
 847             gv2(RC_INT, RC_ECX);
 848             r = vtop[-1].r;
 849             o(0xd3); /* shl/shr/sar %cl, r */
 850             o(opc | r);
 851         }
 852         vtop--;
 853         break;
 854     case '/':
 855     case TOK_UDIV:
 856     case TOK_PDIV:
 857     case '%':
 858     case TOK_UMOD:
 859     case TOK_UMULL:
 860         /* first operand must be in eax */
 861         /* XXX: need better constraint for second operand */
 862         gv2(RC_EAX, RC_ECX);
 863         r = vtop[-1].r;
 864         fr = vtop[0].r;
 865         vtop--;
 866         save_reg(TREG_EDX);
 867         /* save EAX too if used otherwise */
 868         save_reg_upstack(TREG_EAX, 1);
 869         if (op == TOK_UMULL) {
 870             o(0xf7); /* mul fr */
 871             o(0xe0 + fr);
 872             vtop->r2 = TREG_EDX;
 873             r = TREG_EAX;
 874         } else {
 875             if (op == TOK_UDIV || op == TOK_UMOD) {
 876                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 877                 o(0xf0 + fr);
 878             } else {
 879                 o(0xf799); /* cltd, idiv fr, %eax */
 880                 o(0xf8 + fr);
 881             }
 882             if (op == '%' || op == TOK_UMOD)
 883                 r = TREG_EDX;
 884             else
 885                 r = TREG_EAX;
 886         }
 887         vtop->r = r;
 888         break;
 889     default:
 890         opc = 7;
 891         goto gen_op8;
 892     }
 893 }
 894
 895 /* generate a floating point operation 'v = t1 op t2' instruction. The
 896    two operands are guaranteed to have the same floating point type */
 897 /* XXX: need to use ST1 too */
 898 ST_FUNC void gen_opf(int op)
 899 {
 900     int a, ft, fc, swapped, r;
 901
 902     /* convert constants to memory references */
 903     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 904         vswap();
 905         gv(RC_FLOAT);
 906         vswap();
 907     }
 908     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 909         gv(RC_FLOAT);
 910
 911     /* must put at least one value in the floating point register */
 912     if ((vtop[-1].r & VT_LVAL) &&
 913         (vtop[0].r & VT_LVAL)) {
 914         vswap();
 915         gv(RC_FLOAT);
 916         vswap();
 917     }
 918     swapped = 0;
 919     /* swap the stack if needed so that t1 is the register and t2 is
 920        the memory reference */
 921     if (vtop[-1].r & VT_LVAL) {
 922         vswap();
 923         swapped = 1;
 924     }
 925     if (op >= TOK_ULT && op <= TOK_GT) {
 926         /* load on stack second operand */
 927         load(TREG_ST0, vtop);
 928         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 929         if (op == TOK_GE || op == TOK_GT)
 930             swapped = !swapped;
 931         else if (op == TOK_EQ || op == TOK_NE)
 932             swapped = 0;
 933         if (swapped)
 934             o(0xc9d9); /* fxch %st(1) */
 935         if (op == TOK_EQ || op == TOK_NE)
 936             o(0xe9da); /* fucompp */
 937         else
 938             o(0xd9de); /* fcompp */
 939         o(0xe0df); /* fnstsw %ax */
 940         if (op == TOK_EQ) {
 941             o(0x45e480); /* and $0x45, %ah */
 942             o(0x40fC80); /* cmp $0x40, %ah */
 943         } else if (op == TOK_NE) {
 944             o(0x45e480); /* and $0x45, %ah */
 945             o(0x40f480); /* xor $0x40, %ah */
 946             op = TOK_NE;
 947         } else if (op == TOK_GE || op == TOK_LE) {
 948             o(0x05c4f6); /* test $0x05, %ah */
 949             op = TOK_EQ;
 950         } else {
 951             o(0x45c4f6); /* test $0x45, %ah */
 952             op = TOK_EQ;
 953         }
 954         vtop--;
 955         vtop->r = VT_CMP;
 956         vtop->c.i = op;
 957     } else {
 958         /* no memory reference possible for long double operations */
 959         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 960             load(TREG_ST0, vtop);
 961             swapped = !swapped;
 962         }
 963
 964         switch(op) {
 965         default:
 966         case '+':
 967             a = 0;
 968             break;
 969         case '-':
 970             a = 4;
 971             if (swapped)
 972                 a++;
 973             break;
 974         case '*':
 975             a = 1;
 976             break;
 977         case '/':
 978             a = 6;
 979             if (swapped)
 980                 a++;
 981             break;
 982         }
 983         ft = vtop->type.t;
 984         fc = vtop->c.i;
 985         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 986             o(0xde); /* fxxxp %st, %st(1) */
 987             o(0xc1 + (a << 3));
 988         } else {
 989             /* if saved lvalue, then we must reload it */
 990             r = vtop->r;
 991             if ((r & VT_VALMASK) == VT_LLOCAL) {
 992                 SValue v1;
 993                 r = get_reg(RC_INT);
 994                 v1.type.t = VT_INT;
 995                 v1.r = VT_LOCAL | VT_LVAL;
 996                 v1.c.i = fc;
 997                 load(r, &v1);
 998                 fc = 0;
 999             }
1000
1001             if ((ft & VT_BTYPE) == VT_DOUBLE)
1002                 o(0xdc);
1003             else
1004                 o(0xd8);
1005             gen_modrm(a, r, vtop->sym, fc);
1006         }
1007         vtop--;
1008     }
1009 }
1010
1011 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1012    and 'long long' cases. */
1013 ST_FUNC void gen_cvt_itof(int t)
1014 {
1015     save_reg(TREG_ST0);
1016     gv(RC_INT);
1017     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
1018         /* signed long long to float/double/long double (unsigned case
1019            is handled generically) */
1020         o(0x50 + vtop->r2); /* push r2 */
1021         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1022         o(0x242cdf); /* fildll (%esp) */
1023         o(0x08c483); /* add $8, %esp */
1024     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
1025                (VT_INT | VT_UNSIGNED)) {
1026         /* unsigned int to float/double/long double */
1027         o(0x6a); /* push $0 */
1028         g(0x00);
1029         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1030         o(0x242cdf); /* fildll (%esp) */
1031         o(0x08c483); /* add $8, %esp */
1032     } else {
1033         /* int to float/double/long double */
1034         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1035         o(0x2404db); /* fildl (%esp) */
1036         o(0x04c483); /* add $4, %esp */
1037     }
1038     vtop->r = TREG_ST0;
1039 }
1040
1041 /* convert fp to int 't' type */
1042 ST_FUNC void gen_cvt_ftoi(int t)
1043 {
1044     int bt = vtop->type.t & VT_BTYPE;
1045     if (bt == VT_FLOAT)
1046         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1047     else if (bt == VT_LDOUBLE)
1048         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1049     else
1050         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1051     vswap();
1052     gfunc_call(1);
1053     vpushi(0);
1054     vtop->r = REG_IRET;
1055     vtop->r2 = REG_LRET;
1056 }
1057
1058 /* convert from one floating point type to another */
1059 ST_FUNC void gen_cvt_ftof(int t)
1060 {
1061     /* all we have to do on i386 is to put the float in a register */
1062     gv(RC_FLOAT);
1063 }
1064
1065 /* computed goto support */
1066 ST_FUNC void ggoto(void)
1067 {
1068     gcall_or_jmp(1);
1069     vtop--;
1070 }
1071
1072 /* bound check support functions */
1073 #ifdef CONFIG_TCC_BCHECK
1074
1075 /* generate a bounded pointer addition */
1076 ST_FUNC void gen_bounded_ptr_add(void)
1077 {
1078     /* prepare fast i386 function call (args in eax and edx) */
1079     gv2(RC_EAX, RC_EDX);
1080     /* save all temporary registers */
1081     vtop -= 2;
1082     save_regs(0);
1083     /* do a fast function call */
1084     gen_static_call(TOK___bound_ptr_add);
1085     /* returned pointer is in eax */
1086     vtop++;
1087     vtop->r = TREG_EAX | VT_BOUNDED;
1088     /* address of bounding function call point */
1089     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1090 }
1091
1092 /* patch pointer addition in vtop so that pointer dereferencing is
1093    also tested */
1094 ST_FUNC void gen_bounded_ptr_deref(void)
1095 {
1096     addr_t func;
1097     int  size, align;
1098     Elf32_Rel *rel;
1099     Sym *sym;
1100
1101     size = 0;
1102     /* XXX: put that code in generic part of tcc */
1103     if (!is_float(vtop->type.t)) {
1104         if (vtop->r & VT_LVAL_BYTE)
1105             size = 1;
1106         else if (vtop->r & VT_LVAL_SHORT)
1107             size = 2;
1108     }
1109     if (!size)
1110         size = type_size(&vtop->type, &align);
1111     switch(size) {
1112     case  1: func = TOK___bound_ptr_indir1; break;
1113     case  2: func = TOK___bound_ptr_indir2; break;
1114     case  4: func = TOK___bound_ptr_indir4; break;
1115     case  8: func = TOK___bound_ptr_indir8; break;
1116     case 12: func = TOK___bound_ptr_indir12; break;
1117     case 16: func = TOK___bound_ptr_indir16; break;
1118     default:
1119         tcc_error("unhandled size when dereferencing bounded pointer");
1120         func = 0;
1121         break;
1122     }
1123
1124     /* patch relocation */
1125     /* XXX: find a better solution ? */
1126     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1127     sym = external_global_sym(func, &func_old_type, 0);
1128     if (!sym->c)
1129         put_extern_sym(sym, NULL, 0, 0);
1130     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1131 }
1132 #endif
1133
1134 /* Save the stack pointer onto the stack */
1135 ST_FUNC void gen_vla_sp_save(int addr) {
1136     /* mov %esp,addr(%ebp)*/
1137     o(0x89);
1138     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1139 }
1140
1141 /* Restore the SP from a location on the stack */
1142 ST_FUNC void gen_vla_sp_restore(int addr) {
1143     o(0x8b);
1144     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1145 }
1146
1147 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1148 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1149 #ifdef TCC_TARGET_PE
1150     /* alloca does more than just adjust %rsp on Windows */
1151     vpush_global_sym(&func_old_type, TOK_alloca);
1152     vswap(); /* Move alloca ref past allocation size */
1153     gfunc_call(1);
1154 #else
1155     int r;
1156     r = gv(RC_INT); /* allocation size */
1157     /* sub r,%rsp */
1158     o(0x2b);
1159     o(0xe0 | r);
1160     /* We align to 16 bytes rather than align */
1161     /* and ~15, %esp */
1162     o(0xf0e483);
1163     vpop();
1164 #endif
1165 }
1166
1167 /* end of X86 code generator */
1168 /*************************************************************/
1169 #endif
1170 /*************************************************************/