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 unsigned long 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     int n, *ptr;
 142     while (t) {
 143         ptr = (int *)(cur_text_section->data + t);
 144         n = *ptr; /* next value */
 145         *ptr = a - t - 4;
 146         t = n;
 147     }
 148 }
 149
 150 ST_FUNC void gsym(int t)
 151 {
 152     gsym_addr(t, ind);
 153 }
 154
 155 /* psym is used to put an instruction with a data field which is a
 156    reference to a symbol. It is in fact the same as oad ! */
 157 #define psym oad
 158
 159 /* instruction + 4 bytes data. Return the address of the data */
 160 ST_FUNC int oad(int c, int s)
 161 {
 162     int ind1;
 163
 164     o(c);
 165     ind1 = ind + 4;
 166     if (ind1 > cur_text_section->data_allocated)
 167         section_realloc(cur_text_section, ind1);
 168     *(int *)(cur_text_section->data + ind) = s;
 169     s = ind;
 170     ind = ind1;
 171     return s;
 172 }
 173
 174 /* output constant with relocation if 'r & VT_SYM' is true */
 175 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 176 {
 177     if (r & VT_SYM)
 178         greloc(cur_text_section, sym, ind, R_386_32);
 179     gen_le32(c);
 180 }
 181
 182 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 183 {
 184     if (r & VT_SYM)
 185         greloc(cur_text_section, sym, ind, R_386_PC32);
 186     gen_le32(c - 4);
 187 }
 188
 189 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 190    opcode bits */
 191 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 192 {
 193     op_reg = op_reg << 3;
 194     if ((r & VT_VALMASK) == VT_CONST) {
 195         /* constant memory reference */
 196         o(0x05 | op_reg);
 197         gen_addr32(r, sym, c);
 198     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 199         /* currently, we use only ebp as base */
 200         if (c == (char)c) {
 201             /* short reference */
 202             o(0x45 | op_reg);
 203             g(c);
 204         } else {
 205             oad(0x85 | op_reg, c);
 206         }
 207     } else {
 208         g(0x00 | op_reg | (r & VT_VALMASK));
 209     }
 210 }
 211
 212 /* load 'r' from value 'sv' */
 213 ST_FUNC void load(int r, SValue *sv)
 214 {
 215     int v, t, ft, fc, fr;
 216     SValue v1;
 217
 218 #ifdef TCC_TARGET_PE
 219     SValue v2;
 220     sv = pe_getimport(sv, &v2);
 221 #endif
 222
 223     fr = sv->r;
 224     ft = sv->type.t;
 225     fc = sv->c.ul;
 226
 227     v = fr & VT_VALMASK;
 228     if (fr & VT_LVAL) {
 229         if (v == VT_LLOCAL) {
 230             v1.type.t = VT_INT;
 231             v1.r = VT_LOCAL | VT_LVAL;
 232             v1.c.ul = fc;
 233             fr = r;
 234             if (!(reg_classes[fr] & RC_INT))
 235                 fr = get_reg(RC_INT);
 236             load(fr, &v1);
 237         }
 238         if ((ft & VT_BTYPE) == VT_FLOAT) {
 239             o(0xd9); /* flds */
 240             r = 0;
 241         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 242             o(0xdd); /* fldl */
 243             r = 0;
 244         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 245             o(0xdb); /* fldt */
 246             r = 5;
 247         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 248             o(0xbe0f);   /* movsbl */
 249         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 250             o(0xb60f);   /* movzbl */
 251         } else if ((ft & VT_TYPE) == VT_SHORT) {
 252             o(0xbf0f);   /* movswl */
 253         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 254             o(0xb70f);   /* movzwl */
 255         } else {
 256             o(0x8b);     /* movl */
 257         }
 258         gen_modrm(r, fr, sv->sym, fc);
 259     } else {
 260         if (v == VT_CONST) {
 261             o(0xb8 + r); /* mov $xx, r */
 262             gen_addr32(fr, sv->sym, fc);
 263         } else if (v == VT_LOCAL) {
 264             if (fc) {
 265                 o(0x8d); /* lea xxx(%ebp), r */
 266                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 267             } else {
 268                 o(0x89);
 269                 o(0xe8 + r); /* mov %ebp, r */
 270             }
 271         } else if (v == VT_CMP) {
 272             oad(0xb8 + r, 0); /* mov $0, r */
 273             o(0x0f); /* setxx %br */
 274             o(fc);
 275             o(0xc0 + r);
 276         } else if (v == VT_JMP || v == VT_JMPI) {
 277             t = v & 1;
 278             oad(0xb8 + r, t); /* mov $1, r */
 279             o(0x05eb); /* jmp after */
 280             gsym(fc);
 281             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 282         } else if (v != r) {
 283             o(0x89);
 284             o(0xc0 + r + v * 8); /* mov v, r */
 285         }
 286     }
 287 }
 288
 289 /* store register 'r' in lvalue 'v' */
 290 ST_FUNC void store(int r, SValue *v)
 291 {
 292     int fr, bt, ft, fc;
 293
 294 #ifdef TCC_TARGET_PE
 295     SValue v2;
 296     v = pe_getimport(v, &v2);
 297 #endif
 298
 299     ft = v->type.t;
 300     fc = v->c.ul;
 301     fr = v->r & VT_VALMASK;
 302     bt = ft & VT_BTYPE;
 303     /* XXX: incorrect if float reg to reg */
 304     if (bt == VT_FLOAT) {
 305         o(0xd9); /* fsts */
 306         r = 2;
 307     } else if (bt == VT_DOUBLE) {
 308         o(0xdd); /* fstpl */
 309         r = 2;
 310     } else if (bt == VT_LDOUBLE) {
 311         o(0xc0d9); /* fld %st(0) */
 312         o(0xdb); /* fstpt */
 313         r = 7;
 314     } else {
 315         if (bt == VT_SHORT)
 316             o(0x66);
 317         if (bt == VT_BYTE || bt == VT_BOOL)
 318             o(0x88);
 319         else
 320             o(0x89);
 321     }
 322     if (fr == VT_CONST ||
 323         fr == VT_LOCAL ||
 324         (v->r & VT_LVAL)) {
 325         gen_modrm(r, v->r, v->sym, fc);
 326     } else if (fr != r) {
 327         o(0xc0 + fr + r * 8); /* mov r, fr */
 328     }
 329 }
 330
 331 static void gadd_sp(int val)
 332 {
 333     if (val == (char)val) {
 334         o(0xc483);
 335         g(val);
 336     } else {
 337         oad(0xc481, val); /* add $xxx, %esp */
 338     }
 339 }
 340
 341 static void gen_static_call(int v)
 342 {
 343     Sym *sym;
 344
 345     sym = external_global_sym(v, &func_old_type, 0);
 346     oad(0xe8, -4);
 347     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 348 }
 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) {
 355         /* constant case */
 356         if (vtop->r & VT_SYM) {
 357             /* relocation case */
 358             greloc(cur_text_section, vtop->sym,
 359                    ind + 1, R_386_PC32);
 360         } else {
 361             /* put an empty PC32 relocation */
 362             put_elf_reloc(symtab_section, cur_text_section,
 363                           ind + 1, R_386_PC32, 0);
 364         }
 365         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 366     } else {
 367         /* otherwise, indirect call */
 368         r = gv(RC_INT);
 369         o(0xff); /* call/jmp *r */
 370         o(0xd0 + r + (is_jmp << 4));
 371     }
 372 }
 373
 374 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 375 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 376
 377 /* Return the number of registers needed to return the struct, or 0 if
 378    returning via struct pointer. */
 379 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align)
 380 {
 381 #ifdef TCC_TARGET_PE
 382     int size, align;
 383
 384     *ret_align = 1; // Never have to re-align return values for x86
 385     size = type_size(vt, &align);
 386     if (size > 8) {
 387         return 0;
 388     } else if (size > 4) {
 389         ret->ref = NULL;
 390         ret->t = VT_LLONG;
 391         return 1;
 392     } else {
 393         ret->ref = NULL;
 394         ret->t = VT_INT;
 395         return 1;
 396     }
 397 #else
 398     *ret_align = 1; // Never have to re-align return values for x86
 399     return 0;
 400 #endif
 401 }
 402
 403 /* Generate function call. The function address is pushed first, then
 404    all the parameters in call order. This functions pops all the
 405    parameters and the function address. */
 406 ST_FUNC void gfunc_call(int nb_args)
 407 {
 408     int size, align, r, args_size, i, func_call;
 409     Sym *func_sym;
 410
 411     args_size = 0;
 412     for(i = 0;i < nb_args; i++) {
 413         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 414             size = type_size(&vtop->type, &align);
 415             /* align to stack align size */
 416             size = (size + 3) & ~3;
 417             /* allocate the necessary size on stack */
 418             oad(0xec81, size); /* sub $xxx, %esp */
 419             /* generate structure store */
 420             r = get_reg(RC_INT);
 421             o(0x89); /* mov %esp, r */
 422             o(0xe0 + r);
 423             vset(&vtop->type, r | VT_LVAL, 0);
 424             vswap();
 425             vstore();
 426             args_size += size;
 427         } else if (is_float(vtop->type.t)) {
 428             gv(RC_FLOAT); /* only one float register */
 429             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 430                 size = 4;
 431             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 432                 size = 8;
 433             else
 434                 size = 12;
 435             oad(0xec81, size); /* sub $xxx, %esp */
 436             if (size == 12)
 437                 o(0x7cdb);
 438             else
 439                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 440             g(0x24);
 441             g(0x00);
 442             args_size += size;
 443         } else {
 444             /* simple type (currently always same size) */
 445             /* XXX: implicit cast ? */
 446             r = gv(RC_INT);
 447             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 448                 size = 8;
 449                 o(0x50 + vtop->r2); /* push r */
 450             } else {
 451                 size = 4;
 452             }
 453             o(0x50 + r); /* push r */
 454             args_size += size;
 455         }
 456         vtop--;
 457     }
 458     save_regs(0); /* save used temporary registers */
 459     func_sym = vtop->type.ref;
 460     func_call = func_sym->a.func_call;
 461     /* fast call case */
 462     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 463         func_call == FUNC_FASTCALLW) {
 464         int fastcall_nb_regs;
 465         uint8_t *fastcall_regs_ptr;
 466         if (func_call == FUNC_FASTCALLW) {
 467             fastcall_regs_ptr = fastcallw_regs;
 468             fastcall_nb_regs = 2;
 469         } else {
 470             fastcall_regs_ptr = fastcall_regs;
 471             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 472         }
 473         for(i = 0;i < fastcall_nb_regs; i++) {
 474             if (args_size <= 0)
 475                 break;
 476             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 477             /* XXX: incorrect for struct/floats */
 478             args_size -= 4;
 479         }
 480     }
 481 #ifndef TCC_TARGET_PE
 482     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 483         args_size -= 4;
 484 #endif
 485     gcall_or_jmp(0);
 486
 487     if (args_size && func_call != FUNC_STDCALL)
 488         gadd_sp(args_size);
 489     vtop--;
 490 }
 491
 492 #ifdef TCC_TARGET_PE
 493 #define FUNC_PROLOG_SIZE 10
 494 #else
 495 #define FUNC_PROLOG_SIZE 9
 496 #endif
 497
 498 /* generate function prolog of type 't' */
 499 ST_FUNC void gfunc_prolog(CType *func_type)
 500 {
 501     int addr, align, size, func_call, fastcall_nb_regs;
 502     int param_index, param_addr;
 503     uint8_t *fastcall_regs_ptr;
 504     Sym *sym;
 505     CType *type;
 506
 507     sym = func_type->ref;
 508     func_call = sym->a.func_call;
 509     addr = 8;
 510     loc = 0;
 511     func_vc = 0;
 512
 513     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 514         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 515         fastcall_regs_ptr = fastcall_regs;
 516     } else if (func_call == FUNC_FASTCALLW) {
 517         fastcall_nb_regs = 2;
 518         fastcall_regs_ptr = fastcallw_regs;
 519     } else {
 520         fastcall_nb_regs = 0;
 521         fastcall_regs_ptr = NULL;
 522     }
 523     param_index = 0;
 524
 525     ind += FUNC_PROLOG_SIZE;
 526     func_sub_sp_offset = ind;
 527     /* if the function returns a structure, then add an
 528        implicit pointer parameter */
 529     func_vt = sym->type;
 530     func_var = (sym->c == FUNC_ELLIPSIS);
 531 #ifdef TCC_TARGET_PE
 532     size = type_size(&func_vt,&align);
 533     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 534 #else
 535     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 536 #endif
 537         /* XXX: fastcall case ? */
 538         func_vc = addr;
 539         addr += 4;
 540         param_index++;
 541     }
 542     /* define parameters */
 543     while ((sym = sym->next) != NULL) {
 544         type = &sym->type;
 545         size = type_size(type, &align);
 546         size = (size + 3) & ~3;
 547 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 548         /* structs are passed as pointer */
 549         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 550             size = 4;
 551         }
 552 #endif
 553         if (param_index < fastcall_nb_regs) {
 554             /* save FASTCALL register */
 555             loc -= 4;
 556             o(0x89);     /* movl */
 557             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 558             param_addr = loc;
 559         } else {
 560             param_addr = addr;
 561             addr += size;
 562         }
 563         sym_push(sym->v & ~SYM_FIELD, type,
 564                  VT_LOCAL | lvalue_type(type->t), param_addr);
 565         param_index++;
 566     }
 567     func_ret_sub = 0;
 568     /* pascal type call ? */
 569     if (func_call == FUNC_STDCALL)
 570         func_ret_sub = addr - 8;
 571 #ifndef TCC_TARGET_PE
 572     else if (func_vc)
 573         func_ret_sub = 4;
 574 #endif
 575
 576 #ifdef CONFIG_TCC_BCHECK
 577     /* leave some room for bound checking code */
 578     if (tcc_state->do_bounds_check) {
 579         oad(0xb8, 0); /* lbound section pointer */
 580         oad(0xb8, 0); /* call to function */
 581         func_bound_offset = lbounds_section->data_offset;
 582     }
 583 #endif
 584 }
 585
 586 /* generate function epilog */
 587 ST_FUNC void gfunc_epilog(void)
 588 {
 589     int v, saved_ind;
 590
 591 #ifdef CONFIG_TCC_BCHECK
 592     if (tcc_state->do_bounds_check
 593      && func_bound_offset != lbounds_section->data_offset) {
 594         int saved_ind;
 595         int *bounds_ptr;
 596         Sym *sym_data;
 597         /* add end of table info */
 598         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 599         *bounds_ptr = 0;
 600         /* generate bound local allocation */
 601         saved_ind = ind;
 602         ind = func_sub_sp_offset;
 603         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 604                                func_bound_offset, lbounds_section->data_offset);
 605         greloc(cur_text_section, sym_data,
 606                ind + 1, R_386_32);
 607         oad(0xb8, 0); /* mov %eax, xxx */
 608         gen_static_call(TOK___bound_local_new);
 609
 610         ind = saved_ind;
 611         /* generate bound check local freeing */
 612         o(0x5250); /* save returned value, if any */
 613         greloc(cur_text_section, sym_data,
 614                ind + 1, R_386_32);
 615         oad(0xb8, 0); /* mov %eax, xxx */
 616         gen_static_call(TOK___bound_local_delete);
 617
 618         o(0x585a); /* restore returned value, if any */
 619     }
 620 #endif
 621     o(0xc9); /* leave */
 622     if (func_ret_sub == 0) {
 623         o(0xc3); /* ret */
 624     } else {
 625         o(0xc2); /* ret n */
 626         g(func_ret_sub);
 627         g(func_ret_sub >> 8);
 628     }
 629     /* align local size to word & save local variables */
 630
 631     v = (-loc + 3) & -4;
 632     saved_ind = ind;
 633     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 634 #ifdef TCC_TARGET_PE
 635     if (v >= 4096) {
 636         oad(0xb8, v); /* mov stacksize, %eax */
 637         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 638     } else
 639 #endif
 640     {
 641         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 642         o(0xec81);  /* sub esp, stacksize */
 643         gen_le32(v);
 644 #if FUNC_PROLOG_SIZE == 10
 645         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 646 #endif
 647     }
 648     ind = saved_ind;
 649 }
 650
 651 /* generate a jump to a label */
 652 ST_FUNC int gjmp(int t)
 653 {
 654     return psym(0xe9, t);
 655 }
 656
 657 /* generate a jump to a fixed address */
 658 ST_FUNC void gjmp_addr(int a)
 659 {
 660     int r;
 661     r = a - ind - 2;
 662     if (r == (char)r) {
 663         g(0xeb);
 664         g(r);
 665     } else {
 666         oad(0xe9, a - ind - 5);
 667     }
 668 }
 669
 670 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 671 ST_FUNC int gtst(int inv, int t)
 672 {
 673     int v, *p;
 674
 675     v = vtop->r & VT_VALMASK;
 676     if (v == VT_CMP) {
 677         /* fast case : can jump directly since flags are set */
 678         g(0x0f);
 679         t = psym((vtop->c.i - 16) ^ inv, t);
 680     } else if (v == VT_JMP || v == VT_JMPI) {
 681         /* && or || optimization */
 682         if ((v & 1) == inv) {
 683             /* insert vtop->c jump list in t */
 684             p = &vtop->c.i;
 685             while (*p != 0)
 686                 p = (int *)(cur_text_section->data + *p);
 687             *p = t;
 688             t = vtop->c.i;
 689         } else {
 690             t = gjmp(t);
 691             gsym(vtop->c.i);
 692         }
 693     }
 694     vtop--;
 695     return t;
 696 }
 697
 698 /* generate an integer binary operation */
 699 ST_FUNC void gen_opi(int op)
 700 {
 701     int r, fr, opc, c;
 702
 703     switch(op) {
 704     case '+':
 705     case TOK_ADDC1: /* add with carry generation */
 706         opc = 0;
 707     gen_op8:
 708         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 709             /* constant case */
 710             vswap();
 711             r = gv(RC_INT);
 712             vswap();
 713             c = vtop->c.i;
 714             if (c == (char)c) {
 715                 /* generate inc and dec for smaller code */
 716                 if (c==1 && opc==0) {
 717                     o (0x40 | r); // inc
 718                 } else if (c==1 && opc==5) {
 719                     o (0x48 | r); // dec
 720                 } else {
 721                     o(0x83);
 722                     o(0xc0 | (opc << 3) | r);
 723                     g(c);
 724                 }
 725             } else {
 726                 o(0x81);
 727                 oad(0xc0 | (opc << 3) | r, c);
 728             }
 729         } else {
 730             gv2(RC_INT, RC_INT);
 731             r = vtop[-1].r;
 732             fr = vtop[0].r;
 733             o((opc << 3) | 0x01);
 734             o(0xc0 + r + fr * 8);
 735         }
 736         vtop--;
 737         if (op >= TOK_ULT && op <= TOK_GT) {
 738             vtop->r = VT_CMP;
 739             vtop->c.i = op;
 740         }
 741         break;
 742     case '-':
 743     case TOK_SUBC1: /* sub with carry generation */
 744         opc = 5;
 745         goto gen_op8;
 746     case TOK_ADDC2: /* add with carry use */
 747         opc = 2;
 748         goto gen_op8;
 749     case TOK_SUBC2: /* sub with carry use */
 750         opc = 3;
 751         goto gen_op8;
 752     case '&':
 753         opc = 4;
 754         goto gen_op8;
 755     case '^':
 756         opc = 6;
 757         goto gen_op8;
 758     case '|':
 759         opc = 1;
 760         goto gen_op8;
 761     case '*':
 762         gv2(RC_INT, RC_INT);
 763         r = vtop[-1].r;
 764         fr = vtop[0].r;
 765         vtop--;
 766         o(0xaf0f); /* imul fr, r */
 767         o(0xc0 + fr + r * 8);
 768         break;
 769     case TOK_SHL:
 770         opc = 4;
 771         goto gen_shift;
 772     case TOK_SHR:
 773         opc = 5;
 774         goto gen_shift;
 775     case TOK_SAR:
 776         opc = 7;
 777     gen_shift:
 778         opc = 0xc0 | (opc << 3);
 779         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 780             /* constant case */
 781             vswap();
 782             r = gv(RC_INT);
 783             vswap();
 784             c = vtop->c.i & 0x1f;
 785             o(0xc1); /* shl/shr/sar $xxx, r */
 786             o(opc | r);
 787             g(c);
 788         } else {
 789             /* we generate the shift in ecx */
 790             gv2(RC_INT, RC_ECX);
 791             r = vtop[-1].r;
 792             o(0xd3); /* shl/shr/sar %cl, r */
 793             o(opc | r);
 794         }
 795         vtop--;
 796         break;
 797     case '/':
 798     case TOK_UDIV:
 799     case TOK_PDIV:
 800     case '%':
 801     case TOK_UMOD:
 802     case TOK_UMULL:
 803         /* first operand must be in eax */
 804         /* XXX: need better constraint for second operand */
 805         gv2(RC_EAX, RC_ECX);
 806         r = vtop[-1].r;
 807         fr = vtop[0].r;
 808         vtop--;
 809         save_reg(TREG_EDX);
 810         if (op == TOK_UMULL) {
 811             o(0xf7); /* mul fr */
 812             o(0xe0 + fr);
 813             vtop->r2 = TREG_EDX;
 814             r = TREG_EAX;
 815         } else {
 816             if (op == TOK_UDIV || op == TOK_UMOD) {
 817                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 818                 o(0xf0 + fr);
 819             } else {
 820                 o(0xf799); /* cltd, idiv fr, %eax */
 821                 o(0xf8 + fr);
 822             }
 823             if (op == '%' || op == TOK_UMOD)
 824                 r = TREG_EDX;
 825             else
 826                 r = TREG_EAX;
 827         }
 828         vtop->r = r;
 829         break;
 830     default:
 831         opc = 7;
 832         goto gen_op8;
 833     }
 834 }
 835
 836 /* generate a floating point operation 'v = t1 op t2' instruction. The
 837    two operands are guaranted to have the same floating point type */
 838 /* XXX: need to use ST1 too */
 839 ST_FUNC void gen_opf(int op)
 840 {
 841     int a, ft, fc, swapped, r;
 842
 843     /* convert constants to memory references */
 844     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 845         vswap();
 846         gv(RC_FLOAT);
 847         vswap();
 848     }
 849     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 850         gv(RC_FLOAT);
 851
 852     /* must put at least one value in the floating point register */
 853     if ((vtop[-1].r & VT_LVAL) &&
 854         (vtop[0].r & VT_LVAL)) {
 855         vswap();
 856         gv(RC_FLOAT);
 857         vswap();
 858     }
 859     swapped = 0;
 860     /* swap the stack if needed so that t1 is the register and t2 is
 861        the memory reference */
 862     if (vtop[-1].r & VT_LVAL) {
 863         vswap();
 864         swapped = 1;
 865     }
 866     if (op >= TOK_ULT && op <= TOK_GT) {
 867         /* load on stack second operand */
 868         load(TREG_ST0, vtop);
 869         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 870         if (op == TOK_GE || op == TOK_GT)
 871             swapped = !swapped;
 872         else if (op == TOK_EQ || op == TOK_NE)
 873             swapped = 0;
 874         if (swapped)
 875             o(0xc9d9); /* fxch %st(1) */
 876         if (op == TOK_EQ || op == TOK_NE)
 877             o(0xe9da); /* fucompp */
 878         else
 879             o(0xd9de); /* fcompp */
 880         o(0xe0df); /* fnstsw %ax */
 881         if (op == TOK_EQ) {
 882             o(0x45e480); /* and $0x45, %ah */
 883             o(0x40fC80); /* cmp $0x40, %ah */
 884         } else if (op == TOK_NE) {
 885             o(0x45e480); /* and $0x45, %ah */
 886             o(0x40f480); /* xor $0x40, %ah */
 887             op = TOK_NE;
 888         } else if (op == TOK_GE || op == TOK_LE) {
 889             o(0x05c4f6); /* test $0x05, %ah */
 890             op = TOK_EQ;
 891         } else {
 892             o(0x45c4f6); /* test $0x45, %ah */
 893             op = TOK_EQ;
 894         }
 895         vtop--;
 896         vtop->r = VT_CMP;
 897         vtop->c.i = op;
 898     } else {
 899         /* no memory reference possible for long double operations */
 900         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 901             load(TREG_ST0, vtop);
 902             swapped = !swapped;
 903         }
 904
 905         switch(op) {
 906         default:
 907         case '+':
 908             a = 0;
 909             break;
 910         case '-':
 911             a = 4;
 912             if (swapped)
 913                 a++;
 914             break;
 915         case '*':
 916             a = 1;
 917             break;
 918         case '/':
 919             a = 6;
 920             if (swapped)
 921                 a++;
 922             break;
 923         }
 924         ft = vtop->type.t;
 925         fc = vtop->c.ul;
 926         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 927             o(0xde); /* fxxxp %st, %st(1) */
 928             o(0xc1 + (a << 3));
 929         } else {
 930             /* if saved lvalue, then we must reload it */
 931             r = vtop->r;
 932             if ((r & VT_VALMASK) == VT_LLOCAL) {
 933                 SValue v1;
 934                 r = get_reg(RC_INT);
 935                 v1.type.t = VT_INT;
 936                 v1.r = VT_LOCAL | VT_LVAL;
 937                 v1.c.ul = fc;
 938                 load(r, &v1);
 939                 fc = 0;
 940             }
 941
 942             if ((ft & VT_BTYPE) == VT_DOUBLE)
 943                 o(0xdc);
 944             else
 945                 o(0xd8);
 946             gen_modrm(a, r, vtop->sym, fc);
 947         }
 948         vtop--;
 949     }
 950 }
 951
 952 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 953    and 'long long' cases. */
 954 ST_FUNC void gen_cvt_itof(int t)
 955 {
 956     save_reg(TREG_ST0);
 957     gv(RC_INT);
 958     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 959         /* signed long long to float/double/long double (unsigned case
 960            is handled generically) */
 961         o(0x50 + vtop->r2); /* push r2 */
 962         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 963         o(0x242cdf); /* fildll (%esp) */
 964         o(0x08c483); /* add $8, %esp */
 965     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 966                (VT_INT | VT_UNSIGNED)) {
 967         /* unsigned int to float/double/long double */
 968         o(0x6a); /* push $0 */
 969         g(0x00);
 970         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 971         o(0x242cdf); /* fildll (%esp) */
 972         o(0x08c483); /* add $8, %esp */
 973     } else {
 974         /* int to float/double/long double */
 975         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 976         o(0x2404db); /* fildl (%esp) */
 977         o(0x04c483); /* add $4, %esp */
 978     }
 979     vtop->r = TREG_ST0;
 980 }
 981
 982 /* convert fp to int 't' type */
 983 ST_FUNC void gen_cvt_ftoi(int t)
 984 {
 985  #ifndef COMMIT_4ad186c5ef61_IS_FIXED
 986     /* a good version but it takes a more time to execute */
 987     gv(RC_FLOAT);
 988     save_reg(TREG_EAX);
 989     save_reg(TREG_EDX);
 990     gen_static_call(TOK___tcc_cvt_ftol);
 991     vtop->r = TREG_EAX; /* mark reg as used */
 992     if (t == VT_LLONG)
 993         vtop->r2 = TREG_EDX;
 994  #else
 995     /* a new version with a bug: t2a = 44100312 */
 996     /*
 997     #include<stdio.h>
 998     int main() {
 999         int t1 = 176401255;
1000         float f = 0.25f;
1001         int t2a = (int)(t1 * f); // must be 44100313
1002         int t2b = (int)(t1 * (float)0.25f);
1003         printf("t2a=%d t2b=%d \n",t2a,t2b);
1004         return 0;
1005     }
1006     */
1007     int bt = vtop->type.t & VT_BTYPE;
1008     if (bt == VT_FLOAT)
1009         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1010     else if (bt == VT_LDOUBLE)
1011         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1012     else
1013         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1014     vswap();
1015     gfunc_call(1);
1016     vpushi(0);
1017     vtop->r = REG_IRET;
1018     vtop->r2 = REG_LRET;
1019  #endif
1020 }
1021
1022 /* convert from one floating point type to another */
1023 ST_FUNC void gen_cvt_ftof(int t)
1024 {
1025     /* all we have to do on i386 is to put the float in a register */
1026     gv(RC_FLOAT);
1027 }
1028
1029 /* computed goto support */
1030 ST_FUNC void ggoto(void)
1031 {
1032     gcall_or_jmp(1);
1033     vtop--;
1034 }
1035
1036 /* bound check support functions */
1037 #ifdef CONFIG_TCC_BCHECK
1038
1039 /* generate a bounded pointer addition */
1040 ST_FUNC void gen_bounded_ptr_add(void)
1041 {
1042     /* prepare fast i386 function call (args in eax and edx) */
1043     gv2(RC_EAX, RC_EDX);
1044     /* save all temporary registers */
1045     vtop -= 2;
1046     save_regs(0);
1047     /* do a fast function call */
1048     gen_static_call(TOK___bound_ptr_add);
1049     /* returned pointer is in eax */
1050     vtop++;
1051     vtop->r = TREG_EAX | VT_BOUNDED;
1052     /* address of bounding function call point */
1053     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1054 }
1055
1056 /* patch pointer addition in vtop so that pointer dereferencing is
1057    also tested */
1058 ST_FUNC void gen_bounded_ptr_deref(void)
1059 {
1060     int func;
1061     int size, align;
1062     Elf32_Rel *rel;
1063     Sym *sym;
1064
1065     size = 0;
1066     /* XXX: put that code in generic part of tcc */
1067     if (!is_float(vtop->type.t)) {
1068         if (vtop->r & VT_LVAL_BYTE)
1069             size = 1;
1070         else if (vtop->r & VT_LVAL_SHORT)
1071             size = 2;
1072     }
1073     if (!size)
1074         size = type_size(&vtop->type, &align);
1075     switch(size) {
1076     case  1: func = TOK___bound_ptr_indir1; break;
1077     case  2: func = TOK___bound_ptr_indir2; break;
1078     case  4: func = TOK___bound_ptr_indir4; break;
1079     case  8: func = TOK___bound_ptr_indir8; break;
1080     case 12: func = TOK___bound_ptr_indir12; break;
1081     case 16: func = TOK___bound_ptr_indir16; break;
1082     default:
1083         tcc_error("unhandled size when dereferencing bounded pointer");
1084         func = 0;
1085         break;
1086     }
1087
1088     /* patch relocation */
1089     /* XXX: find a better solution ? */
1090     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1091     sym = external_global_sym(func, &func_old_type, 0);
1092     if (!sym->c)
1093         put_extern_sym(sym, NULL, 0, 0);
1094     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1095 }
1096 #endif
1097
1098 /* Save the stack pointer onto the stack */
1099 ST_FUNC void gen_vla_sp_save(int addr) {
1100     /* mov %esp,addr(%ebp)*/
1101     o(0x89);
1102     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1103 }
1104
1105 /* Restore the SP from a location on the stack */
1106 ST_FUNC void gen_vla_sp_restore(int addr) {
1107     o(0x8b);
1108     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1109 }
1110
1111 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1112 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1113 #ifdef TCC_TARGET_PE
1114     /* alloca does more than just adjust %rsp on Windows */
1115     vpush_global_sym(&func_old_type, TOK_alloca);
1116     vswap(); /* Move alloca ref past allocation size */
1117     gfunc_call(1);
1118     vset(type, REG_IRET, 0);
1119 #else
1120     int r;
1121     r = gv(RC_INT); /* allocation size */
1122     /* sub r,%rsp */
1123     o(0x2b);
1124     o(0xe0 | r);
1125     /* We align to 16 bytes rather than align */
1126     /* and ~15, %esp */
1127     o(0xf0e483);
1128     /* mov %esp, r */
1129     o(0x89);
1130     o(0xe0 | r);
1131     vpop();
1132     vset(type, r, 0);
1133 #endif
1134 }
1135
1136 /* end of X86 code generator */
1137 /*************************************************************/
1138 #endif
1139 /*************************************************************/