i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001-2004 Fabrice Bellard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  19  */
  20
  21 #ifdef TARGET_DEFS_ONLY
  22
  23 /* number of available registers */
  24 #define NB_REGS         4
  25 #define NB_ASM_REGS     8
  26
  27 /* a register can belong to several classes. The classes must be
  28    sorted from more general to more precise (see gv2() code which does
  29    assumptions on it). */
  30 #define RC_INT     0x0001 /* generic integer register */
  31 #define RC_FLOAT   0x0002 /* generic float register */
  32 #define RC_EAX     0x0004
  33 #define RC_ST0     0x0008
  34 #define RC_ECX     0x0010
  35 #define RC_EDX     0x0020
  36 #define RC_IRET    RC_EAX /* function return: integer register */
  37 #define RC_LRET    RC_EDX /* function return: second integer register */
  38 #define RC_FRET    RC_ST0 /* function return: float register */
  39
  40 /* pretty names for the registers */
  41 enum {
  42     TREG_EAX = 0,
  43     TREG_ECX,
  44     TREG_EDX,
  45     TREG_ST0,
  46     TREG_ESP = 4
  47 };
  48
  49 /* return registers for function */
  50 #define REG_IRET TREG_EAX /* single word int return register */
  51 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  52 #define REG_FRET TREG_ST0 /* float return register */
  53
  54 /* defined if function parameters must be evaluated in reverse order */
  55 #define INVERT_FUNC_PARAMS
  56
  57 /* defined if structures are passed as pointers. Otherwise structures
  58    are directly pushed on stack. */
  59 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  60
  61 /* pointer size, in bytes */
  62 #define PTR_SIZE 4
  63
  64 /* long double size and alignment, in bytes */
  65 #define LDOUBLE_SIZE  12
  66 #define LDOUBLE_ALIGN 4
  67 /* maximum alignment (for aligned attribute support) */
  68 #define MAX_ALIGN     8
  69
  70
  71 #define psym oad
  72
  73 /******************************************************/
  74 /* ELF defines */
  75
  76 #define EM_TCC_TARGET EM_386
  77
  78 /* relocation type for 32 bit data relocation */
  79 #define R_DATA_32   R_386_32
  80 #define R_DATA_PTR  R_386_32
  81 #define R_JMP_SLOT  R_386_JMP_SLOT
  82 #define R_COPY      R_386_COPY
  83
  84 #define ELF_START_ADDR 0x08048000
  85 #define ELF_PAGE_SIZE  0x1000
  86
  87 /******************************************************/
  88 #else /* ! TARGET_DEFS_ONLY */
  89 /******************************************************/
  90 #include "tcc.h"
  91
  92 ST_DATA const int reg_classes[NB_REGS] = {
  93     /* eax */ RC_INT | RC_EAX,
  94     /* ecx */ RC_INT | RC_ECX,
  95     /* edx */ RC_INT | RC_EDX,
  96     /* st0 */ RC_FLOAT | RC_ST0,
  97 };
  98
  99 static unsigned long func_sub_sp_offset;
 100 static int func_ret_sub;
 101 #ifdef CONFIG_TCC_BCHECK
 102 static addr_t func_bound_offset;
 103 #endif
 104
 105 /* XXX: make it faster ? */
 106 ST_FUNC void g(int c)
 107 {
 108     int ind1;
 109     ind1 = ind + 1;
 110     if (ind1 > cur_text_section->data_allocated)
 111         section_realloc(cur_text_section, ind1);
 112     cur_text_section->data[ind] = c;
 113     ind = ind1;
 114 }
 115
 116 ST_FUNC void o(unsigned int c)
 117 {
 118     while (c) {
 119         g(c);
 120         c = c >> 8;
 121     }
 122 }
 123
 124 ST_FUNC void gen_le16(int v)
 125 {
 126     g(v);
 127     g(v >> 8);
 128 }
 129
 130 ST_FUNC void gen_le32(int c)
 131 {
 132     g(c);
 133     g(c >> 8);
 134     g(c >> 16);
 135     g(c >> 24);
 136 }
 137
 138 /* output a symbol and patch all calls to it */
 139 ST_FUNC void gsym_addr(int t, int a)
 140 {
 141     while (t) {
 142         unsigned char *ptr = cur_text_section->data + t;
 143         uint32_t n = read32le(ptr); /* next value */
 144         write32le(ptr, a - t - 4);
 145         t = n;
 146     }
 147 }
 148
 149 ST_FUNC void gsym(int t)
 150 {
 151     gsym_addr(t, ind);
 152 }
 153
 154 /* psym is used to put an instruction with a data field which is a
 155    reference to a symbol. It is in fact the same as oad ! */
 156 #define psym oad
 157
 158 /* instruction + 4 bytes data. Return the address of the data */
 159 ST_FUNC int oad(int c, int s)
 160 {
 161     int ind1;
 162
 163     o(c);
 164     ind1 = ind + 4;
 165     if (ind1 > cur_text_section->data_allocated)
 166         section_realloc(cur_text_section, ind1);
 167     write32le(cur_text_section->data + ind, s);
 168     s = ind;
 169     ind = ind1;
 170     return s;
 171 }
 172
 173 /* output constant with relocation if 'r & VT_SYM' is true */
 174 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 175 {
 176     if (r & VT_SYM)
 177         greloc(cur_text_section, sym, ind, R_386_32);
 178     gen_le32(c);
 179 }
 180
 181 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 182 {
 183     if (r & VT_SYM)
 184         greloc(cur_text_section, sym, ind, R_386_PC32);
 185     gen_le32(c - 4);
 186 }
 187
 188 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 189    opcode bits */
 190 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 191 {
 192     op_reg = op_reg << 3;
 193     if ((r & VT_VALMASK) == VT_CONST) {
 194         /* constant memory reference */
 195         o(0x05 | op_reg);
 196         gen_addr32(r, sym, c);
 197     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 198         /* currently, we use only ebp as base */
 199         if (c == (char)c) {
 200             /* short reference */
 201             o(0x45 | op_reg);
 202             g(c);
 203         } else {
 204             oad(0x85 | op_reg, c);
 205         }
 206     } else {
 207         g(0x00 | op_reg | (r & VT_VALMASK));
 208     }
 209 }
 210
 211 /* load 'r' from value 'sv' */
 212 ST_FUNC void load(int r, SValue *sv)
 213 {
 214     int v, t, ft, fc, fr;
 215     SValue v1;
 216
 217 #ifdef TCC_TARGET_PE
 218     SValue v2;
 219     sv = pe_getimport(sv, &v2);
 220 #endif
 221
 222     fr = sv->r;
 223     ft = sv->type.t;
 224     fc = sv->c.i;
 225
 226     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 227
 228     v = fr & VT_VALMASK;
 229     if (fr & VT_LVAL) {
 230         if (v == VT_LLOCAL) {
 231             v1.type.t = VT_INT;
 232             v1.r = VT_LOCAL | VT_LVAL;
 233             v1.c.i = fc;
 234             fr = r;
 235             if (!(reg_classes[fr] & RC_INT))
 236                 fr = get_reg(RC_INT);
 237             load(fr, &v1);
 238         }
 239         if ((ft & VT_BTYPE) == VT_FLOAT) {
 240             o(0xd9); /* flds */
 241             r = 0;
 242         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 243             o(0xdd); /* fldl */
 244             r = 0;
 245         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 246             o(0xdb); /* fldt */
 247             r = 5;
 248         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 249             o(0xbe0f);   /* movsbl */
 250         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 251             o(0xb60f);   /* movzbl */
 252         } else if ((ft & VT_TYPE) == VT_SHORT) {
 253             o(0xbf0f);   /* movswl */
 254         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 255             o(0xb70f);   /* movzwl */
 256         } else {
 257             o(0x8b);     /* movl */
 258         }
 259         gen_modrm(r, fr, sv->sym, fc);
 260     } else {
 261         if (v == VT_CONST) {
 262             o(0xb8 + r); /* mov $xx, r */
 263             gen_addr32(fr, sv->sym, fc);
 264         } else if (v == VT_LOCAL) {
 265             if (fc) {
 266                 o(0x8d); /* lea xxx(%ebp), r */
 267                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 268             } else {
 269                 o(0x89);
 270                 o(0xe8 + r); /* mov %ebp, r */
 271             }
 272         } else if (v == VT_CMP) {
 273             oad(0xb8 + r, 0); /* mov $0, r */
 274             o(0x0f); /* setxx %br */
 275             o(fc);
 276             o(0xc0 + r);
 277         } else if (v == VT_JMP || v == VT_JMPI) {
 278             t = v & 1;
 279             oad(0xb8 + r, t); /* mov $1, r */
 280             o(0x05eb); /* jmp after */
 281             gsym(fc);
 282             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 283         } else if (v != r) {
 284             o(0x89);
 285             o(0xc0 + r + v * 8); /* mov v, r */
 286         }
 287     }
 288 }
 289
 290 /* store register 'r' in lvalue 'v' */
 291 ST_FUNC void store(int r, SValue *v)
 292 {
 293     int fr, bt, ft, fc;
 294
 295 #ifdef TCC_TARGET_PE
 296     SValue v2;
 297     v = pe_getimport(v, &v2);
 298 #endif
 299
 300     ft = v->type.t;
 301     fc = v->c.i;
 302     fr = v->r & VT_VALMASK;
 303     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 304     bt = ft & VT_BTYPE;
 305     /* XXX: incorrect if float reg to reg */
 306     if (bt == VT_FLOAT) {
 307         o(0xd9); /* fsts */
 308         r = 2;
 309     } else if (bt == VT_DOUBLE) {
 310         o(0xdd); /* fstpl */
 311         r = 2;
 312     } else if (bt == VT_LDOUBLE) {
 313         o(0xc0d9); /* fld %st(0) */
 314         o(0xdb); /* fstpt */
 315         r = 7;
 316     } else {
 317         if (bt == VT_SHORT)
 318             o(0x66);
 319         if (bt == VT_BYTE || bt == VT_BOOL)
 320             o(0x88);
 321         else
 322             o(0x89);
 323     }
 324     if (fr == VT_CONST ||
 325         fr == VT_LOCAL ||
 326         (v->r & VT_LVAL)) {
 327         gen_modrm(r, v->r, v->sym, fc);
 328     } else if (fr != r) {
 329         o(0xc0 + fr + r * 8); /* mov r, fr */
 330     }
 331 }
 332
 333 static void gadd_sp(int val)
 334 {
 335     if (val == (char)val) {
 336         o(0xc483);
 337         g(val);
 338     } else {
 339         oad(0xc481, val); /* add $xxx, %esp */
 340     }
 341 }
 342
 343 static void gen_static_call(int v)
 344 {
 345     Sym *sym;
 346
 347     sym = external_global_sym(v, &func_old_type, 0);
 348     oad(0xe8, -4);
 349     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 350 }
 351
 352 /* 'is_jmp' is '1' if it is a jump */
 353 static void gcall_or_jmp(int is_jmp)
 354 {
 355     int r;
 356     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 357         /* constant case */
 358         if (vtop->r & VT_SYM) {
 359             /* relocation case */
 360             greloc(cur_text_section, vtop->sym,
 361                    ind + 1, R_386_PC32);
 362         } else {
 363             /* put an empty PC32 relocation */
 364             put_elf_reloc(symtab_section, cur_text_section,
 365                           ind + 1, R_386_PC32, 0);
 366         }
 367         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 368     } else {
 369         /* otherwise, indirect call */
 370         r = gv(RC_INT);
 371         o(0xff); /* call/jmp *r */
 372         o(0xd0 + r + (is_jmp << 4));
 373     }
 374 }
 375
 376 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 377 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 378
 379 /* Return the number of registers needed to return the struct, or 0 if
 380    returning via struct pointer. */
 381 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 382 {
 383 #ifdef TCC_TARGET_PE
 384     int size, align;
 385
 386     *ret_align = 1; // Never have to re-align return values for x86
 387     *regsize = 4;
 388     size = type_size(vt, &align);
 389     if (size > 8) {
 390         return 0;
 391     } else if (size > 4) {
 392         ret->ref = NULL;
 393         ret->t = VT_LLONG;
 394         return 1;
 395     } else {
 396         ret->ref = NULL;
 397         ret->t = VT_INT;
 398         return 1;
 399     }
 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
 414     args_size = 0;
 415     for(i = 0;i < nb_args; i++) {
 416         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 417             size = type_size(&vtop->type, &align);
 418             /* align to stack align size */
 419             size = (size + 3) & ~3;
 420             /* allocate the necessary size on stack */
 421             oad(0xec81, size); /* sub $xxx, %esp */
 422             /* generate structure store */
 423             r = get_reg(RC_INT);
 424             o(0x89); /* mov %esp, r */
 425             o(0xe0 + r);
 426             vset(&vtop->type, r | VT_LVAL, 0);
 427             vswap();
 428             vstore();
 429             args_size += size;
 430         } else if (is_float(vtop->type.t)) {
 431             gv(RC_FLOAT); /* only one float register */
 432             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 433                 size = 4;
 434             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 435                 size = 8;
 436             else
 437                 size = 12;
 438             oad(0xec81, size); /* sub $xxx, %esp */
 439             if (size == 12)
 440                 o(0x7cdb);
 441             else
 442                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 443             g(0x24);
 444             g(0x00);
 445             args_size += size;
 446         } else {
 447             /* simple type (currently always same size) */
 448             /* XXX: implicit cast ? */
 449             r = gv(RC_INT);
 450             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 451                 size = 8;
 452                 o(0x50 + vtop->r2); /* push r */
 453             } else {
 454                 size = 4;
 455             }
 456             o(0x50 + r); /* push r */
 457             args_size += size;
 458         }
 459         vtop--;
 460     }
 461     save_regs(0); /* save used temporary registers */
 462     func_sym = vtop->type.ref;
 463     func_call = func_sym->a.func_call;
 464     /* fast call case */
 465     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 466         func_call == FUNC_FASTCALLW) {
 467         int fastcall_nb_regs;
 468         uint8_t *fastcall_regs_ptr;
 469         if (func_call == FUNC_FASTCALLW) {
 470             fastcall_regs_ptr = fastcallw_regs;
 471             fastcall_nb_regs = 2;
 472         } else {
 473             fastcall_regs_ptr = fastcall_regs;
 474             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 475         }
 476         for(i = 0;i < fastcall_nb_regs; i++) {
 477             if (args_size <= 0)
 478                 break;
 479             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 480             /* XXX: incorrect for struct/floats */
 481             args_size -= 4;
 482         }
 483     }
 484 #ifndef TCC_TARGET_PE
 485     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 486         args_size -= 4;
 487 #endif
 488     gcall_or_jmp(0);
 489
 490     if (args_size && func_call != FUNC_STDCALL)
 491         gadd_sp(args_size);
 492     vtop--;
 493 }
 494
 495 #ifdef TCC_TARGET_PE
 496 #define FUNC_PROLOG_SIZE 10
 497 #else
 498 #define FUNC_PROLOG_SIZE 9
 499 #endif
 500
 501 /* generate function prolog of type 't' */
 502 ST_FUNC void gfunc_prolog(CType *func_type)
 503 {
 504     int addr, align, size, func_call, fastcall_nb_regs;
 505     int param_index, param_addr;
 506     uint8_t *fastcall_regs_ptr;
 507     Sym *sym;
 508     CType *type;
 509
 510     sym = func_type->ref;
 511     func_call = sym->a.func_call;
 512     addr = 8;
 513     loc = 0;
 514     func_vc = 0;
 515
 516     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 517         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 518         fastcall_regs_ptr = fastcall_regs;
 519     } else if (func_call == FUNC_FASTCALLW) {
 520         fastcall_nb_regs = 2;
 521         fastcall_regs_ptr = fastcallw_regs;
 522     } else {
 523         fastcall_nb_regs = 0;
 524         fastcall_regs_ptr = NULL;
 525     }
 526     param_index = 0;
 527
 528     ind += FUNC_PROLOG_SIZE;
 529     func_sub_sp_offset = ind;
 530     /* if the function returns a structure, then add an
 531        implicit pointer parameter */
 532     func_vt = sym->type;
 533     func_var = (sym->c == FUNC_ELLIPSIS);
 534 #ifdef TCC_TARGET_PE
 535     size = type_size(&func_vt,&align);
 536     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 537 #else
 538     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 539 #endif
 540         /* XXX: fastcall case ? */
 541         func_vc = addr;
 542         addr += 4;
 543         param_index++;
 544     }
 545     /* define parameters */
 546     while ((sym = sym->next) != NULL) {
 547         type = &sym->type;
 548         size = type_size(type, &align);
 549         size = (size + 3) & ~3;
 550 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 551         /* structs are passed as pointer */
 552         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 553             size = 4;
 554         }
 555 #endif
 556         if (param_index < fastcall_nb_regs) {
 557             /* save FASTCALL register */
 558             loc -= 4;
 559             o(0x89);     /* movl */
 560             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 561             param_addr = loc;
 562         } else {
 563             param_addr = addr;
 564             addr += size;
 565         }
 566         sym_push(sym->v & ~SYM_FIELD, type,
 567                  VT_LOCAL | lvalue_type(type->t), param_addr);
 568         param_index++;
 569     }
 570     func_ret_sub = 0;
 571     /* pascal type call ? */
 572     if (func_call == FUNC_STDCALL)
 573         func_ret_sub = addr - 8;
 574 #ifndef TCC_TARGET_PE
 575     else if (func_vc)
 576         func_ret_sub = 4;
 577 #endif
 578
 579 #ifdef CONFIG_TCC_BCHECK
 580     /* leave some room for bound checking code */
 581     if (tcc_state->do_bounds_check) {
 582         oad(0xb8, 0); /* lbound section pointer */
 583         oad(0xb8, 0); /* call to function */
 584         func_bound_offset = lbounds_section->data_offset;
 585     }
 586 #endif
 587 }
 588
 589 /* generate function epilog */
 590 ST_FUNC void gfunc_epilog(void)
 591 {
 592     addr_t v, saved_ind;
 593
 594 #ifdef CONFIG_TCC_BCHECK
 595     if (tcc_state->do_bounds_check
 596      && func_bound_offset != lbounds_section->data_offset) {
 597         addr_t saved_ind;
 598         addr_t *bounds_ptr;
 599         Sym *sym_data;
 600         /* add end of table info */
 601         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 602         *bounds_ptr = 0;
 603         /* generate bound local allocation */
 604         saved_ind = ind;
 605         ind = func_sub_sp_offset;
 606         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 607                                func_bound_offset, lbounds_section->data_offset);
 608         greloc(cur_text_section, sym_data,
 609                ind + 1, R_386_32);
 610         oad(0xb8, 0); /* mov %eax, xxx */
 611         gen_static_call(TOK___bound_local_new);
 612
 613         ind = saved_ind;
 614         /* generate bound check local freeing */
 615         o(0x5250); /* save returned value, if any */
 616         greloc(cur_text_section, sym_data,
 617                ind + 1, R_386_32);
 618         oad(0xb8, 0); /* mov %eax, xxx */
 619         gen_static_call(TOK___bound_local_delete);
 620
 621         o(0x585a); /* restore returned value, if any */
 622     }
 623 #endif
 624     o(0xc9); /* leave */
 625     if (func_ret_sub == 0) {
 626         o(0xc3); /* ret */
 627     } else {
 628         o(0xc2); /* ret n */
 629         g(func_ret_sub);
 630         g(func_ret_sub >> 8);
 631     }
 632     /* align local size to word & save local variables */
 633
 634     v = (-loc + 3) & -4;
 635     saved_ind = ind;
 636     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 637 #ifdef TCC_TARGET_PE
 638     if (v >= 4096) {
 639         oad(0xb8, v); /* mov stacksize, %eax */
 640         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 641     } else
 642 #endif
 643     {
 644         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 645         o(0xec81);  /* sub esp, stacksize */
 646         gen_le32(v);
 647 #if FUNC_PROLOG_SIZE == 10
 648         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 649 #endif
 650     }
 651     ind = saved_ind;
 652 }
 653
 654 /* generate a jump to a label */
 655 ST_FUNC int gjmp(int t)
 656 {
 657     return psym(0xe9, t);
 658 }
 659
 660 /* generate a jump to a fixed address */
 661 ST_FUNC void gjmp_addr(int a)
 662 {
 663     int r;
 664     r = a - ind - 2;
 665     if (r == (char)r) {
 666         g(0xeb);
 667         g(r);
 668     } else {
 669         oad(0xe9, a - ind - 5);
 670     }
 671 }
 672
 673 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 674 ST_FUNC int gtst(int inv, int t)
 675 {
 676     int v = vtop->r & VT_VALMASK;
 677     if (v == VT_CMP) {
 678         /* fast case : can jump directly since flags are set */
 679         g(0x0f);
 680         t = psym((vtop->c.i - 16) ^ inv, t);
 681     } else if (v == VT_JMP || v == VT_JMPI) {
 682         /* && or || optimization */
 683         if ((v & 1) == inv) {
 684             /* insert vtop->c jump list in t */
 685             uint32_t n1, n = vtop->c.i;
 686             if (n) {
 687                 while ((n1 = read32le(cur_text_section->data + n)))
 688                     n = n1;
 689                 write32le(cur_text_section->data + n, t);
 690                 t = vtop->c.i;
 691             }
 692         } else {
 693             t = gjmp(t);
 694             gsym(vtop->c.i);
 695         }
 696     }
 697     vtop--;
 698     return t;
 699 }
 700
 701 /* generate an integer binary operation */
 702 ST_FUNC void gen_opi(int op)
 703 {
 704     int r, fr, opc, c;
 705
 706     switch(op) {
 707     case '+':
 708     case TOK_ADDC1: /* add with carry generation */
 709         opc = 0;
 710     gen_op8:
 711         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 712             /* constant case */
 713             vswap();
 714             r = gv(RC_INT);
 715             vswap();
 716             c = vtop->c.i;
 717             if (c == (char)c) {
 718                 /* generate inc and dec for smaller code */
 719                 if (c==1 && opc==0) {
 720                     o (0x40 | r); // inc
 721                 } else if (c==1 && opc==5) {
 722                     o (0x48 | r); // dec
 723                 } else {
 724                     o(0x83);
 725                     o(0xc0 | (opc << 3) | r);
 726                     g(c);
 727                 }
 728             } else {
 729                 o(0x81);
 730                 oad(0xc0 | (opc << 3) | r, c);
 731             }
 732         } else {
 733             gv2(RC_INT, RC_INT);
 734             r = vtop[-1].r;
 735             fr = vtop[0].r;
 736             o((opc << 3) | 0x01);
 737             o(0xc0 + r + fr * 8);
 738         }
 739         vtop--;
 740         if (op >= TOK_ULT && op <= TOK_GT) {
 741             vtop->r = VT_CMP;
 742             vtop->c.i = op;
 743         }
 744         break;
 745     case '-':
 746     case TOK_SUBC1: /* sub with carry generation */
 747         opc = 5;
 748         goto gen_op8;
 749     case TOK_ADDC2: /* add with carry use */
 750         opc = 2;
 751         goto gen_op8;
 752     case TOK_SUBC2: /* sub with carry use */
 753         opc = 3;
 754         goto gen_op8;
 755     case '&':
 756         opc = 4;
 757         goto gen_op8;
 758     case '^':
 759         opc = 6;
 760         goto gen_op8;
 761     case '|':
 762         opc = 1;
 763         goto gen_op8;
 764     case '*':
 765         gv2(RC_INT, RC_INT);
 766         r = vtop[-1].r;
 767         fr = vtop[0].r;
 768         vtop--;
 769         o(0xaf0f); /* imul fr, r */
 770         o(0xc0 + fr + r * 8);
 771         break;
 772     case TOK_SHL:
 773         opc = 4;
 774         goto gen_shift;
 775     case TOK_SHR:
 776         opc = 5;
 777         goto gen_shift;
 778     case TOK_SAR:
 779         opc = 7;
 780     gen_shift:
 781         opc = 0xc0 | (opc << 3);
 782         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 783             /* constant case */
 784             vswap();
 785             r = gv(RC_INT);
 786             vswap();
 787             c = vtop->c.i & 0x1f;
 788             o(0xc1); /* shl/shr/sar $xxx, r */
 789             o(opc | r);
 790             g(c);
 791         } else {
 792             /* we generate the shift in ecx */
 793             gv2(RC_INT, RC_ECX);
 794             r = vtop[-1].r;
 795             o(0xd3); /* shl/shr/sar %cl, r */
 796             o(opc | r);
 797         }
 798         vtop--;
 799         break;
 800     case '/':
 801     case TOK_UDIV:
 802     case TOK_PDIV:
 803     case '%':
 804     case TOK_UMOD:
 805     case TOK_UMULL:
 806         /* first operand must be in eax */
 807         /* XXX: need better constraint for second operand */
 808         gv2(RC_EAX, RC_ECX);
 809         r = vtop[-1].r;
 810         fr = vtop[0].r;
 811         vtop--;
 812         save_reg(TREG_EDX);
 813         if (op == TOK_UMULL) {
 814             o(0xf7); /* mul fr */
 815             o(0xe0 + fr);
 816             vtop->r2 = TREG_EDX;
 817             r = TREG_EAX;
 818         } else {
 819             if (op == TOK_UDIV || op == TOK_UMOD) {
 820                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 821                 o(0xf0 + fr);
 822             } else {
 823                 o(0xf799); /* cltd, idiv fr, %eax */
 824                 o(0xf8 + fr);
 825             }
 826             if (op == '%' || op == TOK_UMOD)
 827                 r = TREG_EDX;
 828             else
 829                 r = TREG_EAX;
 830         }
 831         vtop->r = r;
 832         break;
 833     default:
 834         opc = 7;
 835         goto gen_op8;
 836     }
 837 }
 838
 839 /* generate a floating point operation 'v = t1 op t2' instruction. The
 840    two operands are guaranted to have the same floating point type */
 841 /* XXX: need to use ST1 too */
 842 ST_FUNC void gen_opf(int op)
 843 {
 844     int a, ft, fc, swapped, r;
 845
 846     /* convert constants to memory references */
 847     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 848         vswap();
 849         gv(RC_FLOAT);
 850         vswap();
 851     }
 852     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 853         gv(RC_FLOAT);
 854
 855     /* must put at least one value in the floating point register */
 856     if ((vtop[-1].r & VT_LVAL) &&
 857         (vtop[0].r & VT_LVAL)) {
 858         vswap();
 859         gv(RC_FLOAT);
 860         vswap();
 861     }
 862     swapped = 0;
 863     /* swap the stack if needed so that t1 is the register and t2 is
 864        the memory reference */
 865     if (vtop[-1].r & VT_LVAL) {
 866         vswap();
 867         swapped = 1;
 868     }
 869     if (op >= TOK_ULT && op <= TOK_GT) {
 870         /* load on stack second operand */
 871         load(TREG_ST0, vtop);
 872         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 873         if (op == TOK_GE || op == TOK_GT)
 874             swapped = !swapped;
 875         else if (op == TOK_EQ || op == TOK_NE)
 876             swapped = 0;
 877         if (swapped)
 878             o(0xc9d9); /* fxch %st(1) */
 879         if (op == TOK_EQ || op == TOK_NE)
 880             o(0xe9da); /* fucompp */
 881         else
 882             o(0xd9de); /* fcompp */
 883         o(0xe0df); /* fnstsw %ax */
 884         if (op == TOK_EQ) {
 885             o(0x45e480); /* and $0x45, %ah */
 886             o(0x40fC80); /* cmp $0x40, %ah */
 887         } else if (op == TOK_NE) {
 888             o(0x45e480); /* and $0x45, %ah */
 889             o(0x40f480); /* xor $0x40, %ah */
 890             op = TOK_NE;
 891         } else if (op == TOK_GE || op == TOK_LE) {
 892             o(0x05c4f6); /* test $0x05, %ah */
 893             op = TOK_EQ;
 894         } else {
 895             o(0x45c4f6); /* test $0x45, %ah */
 896             op = TOK_EQ;
 897         }
 898         vtop--;
 899         vtop->r = VT_CMP;
 900         vtop->c.i = op;
 901     } else {
 902         /* no memory reference possible for long double operations */
 903         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 904             load(TREG_ST0, vtop);
 905             swapped = !swapped;
 906         }
 907
 908         switch(op) {
 909         default:
 910         case '+':
 911             a = 0;
 912             break;
 913         case '-':
 914             a = 4;
 915             if (swapped)
 916                 a++;
 917             break;
 918         case '*':
 919             a = 1;
 920             break;
 921         case '/':
 922             a = 6;
 923             if (swapped)
 924                 a++;
 925             break;
 926         }
 927         ft = vtop->type.t;
 928         fc = vtop->c.i;
 929         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 930             o(0xde); /* fxxxp %st, %st(1) */
 931             o(0xc1 + (a << 3));
 932         } else {
 933             /* if saved lvalue, then we must reload it */
 934             r = vtop->r;
 935             if ((r & VT_VALMASK) == VT_LLOCAL) {
 936                 SValue v1;
 937                 r = get_reg(RC_INT);
 938                 v1.type.t = VT_INT;
 939                 v1.r = VT_LOCAL | VT_LVAL;
 940                 v1.c.i = fc;
 941                 load(r, &v1);
 942                 fc = 0;
 943             }
 944
 945             if ((ft & VT_BTYPE) == VT_DOUBLE)
 946                 o(0xdc);
 947             else
 948                 o(0xd8);
 949             gen_modrm(a, r, vtop->sym, fc);
 950         }
 951         vtop--;
 952     }
 953 }
 954
 955 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 956    and 'long long' cases. */
 957 ST_FUNC void gen_cvt_itof(int t)
 958 {
 959     save_reg(TREG_ST0);
 960     gv(RC_INT);
 961     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 962         /* signed long long to float/double/long double (unsigned case
 963            is handled generically) */
 964         o(0x50 + vtop->r2); /* push r2 */
 965         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 966         o(0x242cdf); /* fildll (%esp) */
 967         o(0x08c483); /* add $8, %esp */
 968     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 969                (VT_INT | VT_UNSIGNED)) {
 970         /* unsigned int to float/double/long double */
 971         o(0x6a); /* push $0 */
 972         g(0x00);
 973         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 974         o(0x242cdf); /* fildll (%esp) */
 975         o(0x08c483); /* add $8, %esp */
 976     } else {
 977         /* int to float/double/long double */
 978         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 979         o(0x2404db); /* fildl (%esp) */
 980         o(0x04c483); /* add $4, %esp */
 981     }
 982     vtop->r = TREG_ST0;
 983 }
 984
 985 /* convert fp to int 't' type */
 986 ST_FUNC void gen_cvt_ftoi(int t)
 987 {
 988  #ifndef COMMIT_4ad186c5ef61_IS_FIXED
 989     /* a good version but it takes a more time to execute */
 990     gv(RC_FLOAT);
 991     save_reg(TREG_EAX);
 992     save_reg(TREG_EDX);
 993     gen_static_call(TOK___tcc_cvt_ftol);
 994     vtop->r = TREG_EAX; /* mark reg as used */
 995     if (t == VT_LLONG)
 996         vtop->r2 = TREG_EDX;
 997  #else
 998     /* a new version with a bug: t2a = 44100312 */
 999     /*
1000     #include<stdio.h>
1001     int main() {
1002         int t1 = 176401255;
1003         float f = 0.25f;
1004         int t2a = (int)(t1 * f); // must be 44100313
1005         int t2b = (int)(t1 * (float)0.25f);
1006         printf("t2a=%d t2b=%d \n",t2a,t2b);
1007         return 0;
1008     }
1009     */
1010     int bt = vtop->type.t & VT_BTYPE;
1011     if (bt == VT_FLOAT)
1012         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1013     else if (bt == VT_LDOUBLE)
1014         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1015     else
1016         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1017     vswap();
1018     gfunc_call(1);
1019     vpushi(0);
1020     vtop->r = REG_IRET;
1021     vtop->r2 = REG_LRET;
1022  #endif
1023 }
1024
1025 /* convert from one floating point type to another */
1026 ST_FUNC void gen_cvt_ftof(int t)
1027 {
1028     /* all we have to do on i386 is to put the float in a register */
1029     gv(RC_FLOAT);
1030 }
1031
1032 /* computed goto support */
1033 ST_FUNC void ggoto(void)
1034 {
1035     gcall_or_jmp(1);
1036     vtop--;
1037 }
1038
1039 /* bound check support functions */
1040 #ifdef CONFIG_TCC_BCHECK
1041
1042 /* generate a bounded pointer addition */
1043 ST_FUNC void gen_bounded_ptr_add(void)
1044 {
1045     /* prepare fast i386 function call (args in eax and edx) */
1046     gv2(RC_EAX, RC_EDX);
1047     /* save all temporary registers */
1048     vtop -= 2;
1049     save_regs(0);
1050     /* do a fast function call */
1051     gen_static_call(TOK___bound_ptr_add);
1052     /* returned pointer is in eax */
1053     vtop++;
1054     vtop->r = TREG_EAX | VT_BOUNDED;
1055     /* address of bounding function call point */
1056     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1057 }
1058
1059 /* patch pointer addition in vtop so that pointer dereferencing is
1060    also tested */
1061 ST_FUNC void gen_bounded_ptr_deref(void)
1062 {
1063     addr_t func;
1064     addr_t size, align;
1065     Elf32_Rel *rel;
1066     Sym *sym;
1067
1068     size = 0;
1069     /* XXX: put that code in generic part of tcc */
1070     if (!is_float(vtop->type.t)) {
1071         if (vtop->r & VT_LVAL_BYTE)
1072             size = 1;
1073         else if (vtop->r & VT_LVAL_SHORT)
1074             size = 2;
1075     }
1076     if (!size)
1077         size = type_size(&vtop->type, &align);
1078     switch(size) {
1079     case  1: func = TOK___bound_ptr_indir1; break;
1080     case  2: func = TOK___bound_ptr_indir2; break;
1081     case  4: func = TOK___bound_ptr_indir4; break;
1082     case  8: func = TOK___bound_ptr_indir8; break;
1083     case 12: func = TOK___bound_ptr_indir12; break;
1084     case 16: func = TOK___bound_ptr_indir16; break;
1085     default:
1086         tcc_error("unhandled size when dereferencing bounded pointer");
1087         func = 0;
1088         break;
1089     }
1090
1091     /* patch relocation */
1092     /* XXX: find a better solution ? */
1093     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1094     sym = external_global_sym(func, &func_old_type, 0);
1095     if (!sym->c)
1096         put_extern_sym(sym, NULL, 0, 0);
1097     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1098 }
1099 #endif
1100
1101 /* Save the stack pointer onto the stack */
1102 ST_FUNC void gen_vla_sp_save(int addr) {
1103     /* mov %esp,addr(%ebp)*/
1104     o(0x89);
1105     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1106 }
1107
1108 /* Restore the SP from a location on the stack */
1109 ST_FUNC void gen_vla_sp_restore(int addr) {
1110     o(0x8b);
1111     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1112 }
1113
1114 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1115 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1116 #ifdef TCC_TARGET_PE
1117     /* alloca does more than just adjust %rsp on Windows */
1118     vpush_global_sym(&func_old_type, TOK_alloca);
1119     vswap(); /* Move alloca ref past allocation size */
1120     gfunc_call(1);
1121 #else
1122     int r;
1123     r = gv(RC_INT); /* allocation size */
1124     /* sub r,%rsp */
1125     o(0x2b);
1126     o(0xe0 | r);
1127     /* We align to 16 bytes rather than align */
1128     /* and ~15, %esp */
1129     o(0xf0e483);
1130     vpop();
1131 #endif
1132 }
1133
1134 /* end of X86 code generator */
1135 /*************************************************************/
1136 #endif
1137 /*************************************************************/