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 1 if this function returns via an sret pointer, 0 otherwise */
 378 ST_FUNC int gfunc_sret(CType *vt, CType *ret, int *ret_align)
 379 {
 380 #ifdef TCC_TARGET_PE
 381     int size, align;
 382
 383     *ret_align = 1; // Never have to re-align return values for x86
 384     size = type_size(vt, &align);
 385     if (size > 8) {
 386         return 1;
 387     } else if (size > 4) {
 388         ret->ref = NULL;
 389         ret->t = VT_LLONG;
 390         return 0;
 391     } else {
 392         ret->ref = NULL;
 393         ret->t = VT_INT;
 394         return 0;
 395     }
 396 #else
 397     *ret_align = 1; // Never have to re-align return values for x86
 398     return 1;
 399 #endif
 400 }
 401
 402 /* Generate function call. The function address is pushed first, then
 403    all the parameters in call order. This functions pops all the
 404    parameters and the function address. */
 405 ST_FUNC void gfunc_call(int nb_args)
 406 {
 407     int size, align, r, args_size, i, func_call;
 408     Sym *func_sym;
 409
 410     args_size = 0;
 411     for(i = 0;i < nb_args; i++) {
 412         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 413             size = type_size(&vtop->type, &align);
 414             /* align to stack align size */
 415             size = (size + 3) & ~3;
 416             /* allocate the necessary size on stack */
 417             oad(0xec81, size); /* sub $xxx, %esp */
 418             /* generate structure store */
 419             r = get_reg(RC_INT);
 420             o(0x89); /* mov %esp, r */
 421             o(0xe0 + r);
 422             vset(&vtop->type, r | VT_LVAL, 0);
 423             vswap();
 424             vstore();
 425             args_size += size;
 426         } else if (is_float(vtop->type.t)) {
 427             gv(RC_FLOAT); /* only one float register */
 428             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 429                 size = 4;
 430             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 431                 size = 8;
 432             else
 433                 size = 12;
 434             oad(0xec81, size); /* sub $xxx, %esp */
 435             if (size == 12)
 436                 o(0x7cdb);
 437             else
 438                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 439             g(0x24);
 440             g(0x00);
 441             args_size += size;
 442         } else {
 443             /* simple type (currently always same size) */
 444             /* XXX: implicit cast ? */
 445             r = gv(RC_INT);
 446             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 447                 size = 8;
 448                 o(0x50 + vtop->r2); /* push r */
 449             } else {
 450                 size = 4;
 451             }
 452             o(0x50 + r); /* push r */
 453             args_size += size;
 454         }
 455         vtop--;
 456     }
 457     save_regs(0); /* save used temporary registers */
 458     func_sym = vtop->type.ref;
 459     func_call = FUNC_CALL(func_sym->r);
 460     /* fast call case */
 461     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 462         func_call == FUNC_FASTCALLW) {
 463         int fastcall_nb_regs;
 464         uint8_t *fastcall_regs_ptr;
 465         if (func_call == FUNC_FASTCALLW) {
 466             fastcall_regs_ptr = fastcallw_regs;
 467             fastcall_nb_regs = 2;
 468         } else {
 469             fastcall_regs_ptr = fastcall_regs;
 470             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 471         }
 472         for(i = 0;i < fastcall_nb_regs; i++) {
 473             if (args_size <= 0)
 474                 break;
 475             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 476             /* XXX: incorrect for struct/floats */
 477             args_size -= 4;
 478         }
 479     }
 480 #ifndef TCC_TARGET_PE
 481     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 482         args_size -= 4;
 483 #endif
 484     gcall_or_jmp(0);
 485
 486     if (args_size && func_call != FUNC_STDCALL)
 487         gadd_sp(args_size);
 488     vtop--;
 489 }
 490
 491 #ifdef TCC_TARGET_PE
 492 #define FUNC_PROLOG_SIZE 10
 493 #else
 494 #define FUNC_PROLOG_SIZE 9
 495 #endif
 496
 497 /* generate function prolog of type 't' */
 498 ST_FUNC void gfunc_prolog(CType *func_type)
 499 {
 500     int addr, align, size, func_call, fastcall_nb_regs;
 501     int param_index, param_addr;
 502     uint8_t *fastcall_regs_ptr;
 503     Sym *sym;
 504     CType *type;
 505
 506     sym = func_type->ref;
 507     func_call = FUNC_CALL(sym->r);
 508     addr = 8;
 509     loc = 0;
 510     func_vc = 0;
 511
 512     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 513         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 514         fastcall_regs_ptr = fastcall_regs;
 515     } else if (func_call == FUNC_FASTCALLW) {
 516         fastcall_nb_regs = 2;
 517         fastcall_regs_ptr = fastcallw_regs;
 518     } else {
 519         fastcall_nb_regs = 0;
 520         fastcall_regs_ptr = NULL;
 521     }
 522     param_index = 0;
 523
 524     ind += FUNC_PROLOG_SIZE;
 525     func_sub_sp_offset = ind;
 526     /* if the function returns a structure, then add an
 527        implicit pointer parameter */
 528     func_vt = sym->type;
 529 #ifdef TCC_TARGET_PE
 530     size = type_size(&func_vt,&align);
 531     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 532 #else
 533     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 534 #endif
 535         /* XXX: fastcall case ? */
 536         func_vc = addr;
 537         addr += 4;
 538         param_index++;
 539     }
 540     /* define parameters */
 541     while ((sym = sym->next) != NULL) {
 542         type = &sym->type;
 543         size = type_size(type, &align);
 544         size = (size + 3) & ~3;
 545 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 546         /* structs are passed as pointer */
 547         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 548             size = 4;
 549         }
 550 #endif
 551         if (param_index < fastcall_nb_regs) {
 552             /* save FASTCALL register */
 553             loc -= 4;
 554             o(0x89);     /* movl */
 555             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 556             param_addr = loc;
 557         } else {
 558             param_addr = addr;
 559             addr += size;
 560         }
 561         sym_push(sym->v & ~SYM_FIELD, type,
 562                  VT_LOCAL | lvalue_type(type->t), param_addr);
 563         param_index++;
 564     }
 565     func_ret_sub = 0;
 566     /* pascal type call ? */
 567     if (func_call == FUNC_STDCALL)
 568         func_ret_sub = addr - 8;
 569 #ifndef TCC_TARGET_PE
 570     else if (func_vc)
 571         func_ret_sub = 4;
 572 #endif
 573
 574 #ifdef CONFIG_TCC_BCHECK
 575     /* leave some room for bound checking code */
 576     if (tcc_state->do_bounds_check) {
 577         oad(0xb8, 0); /* lbound section pointer */
 578         oad(0xb8, 0); /* call to function */
 579         func_bound_offset = lbounds_section->data_offset;
 580     }
 581 #endif
 582
 583 #ifndef TCC_TARGET_PE
 584     if (0 == strcmp(funcname, "main"))
 585         gen_static_call(TOK___tcc_fpinit);
 586 #endif
 587
 588 }
 589
 590 /* generate function epilog */
 591 ST_FUNC void gfunc_epilog(void)
 592 {
 593     int v, saved_ind;
 594
 595 #ifdef CONFIG_TCC_BCHECK
 596     if (tcc_state->do_bounds_check
 597      && func_bound_offset != lbounds_section->data_offset) {
 598         int saved_ind;
 599         int *bounds_ptr;
 600         Sym *sym_data;
 601         /* add end of table info */
 602         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 603         *bounds_ptr = 0;
 604         /* generate bound local allocation */
 605         saved_ind = ind;
 606         ind = func_sub_sp_offset;
 607         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 608                                func_bound_offset, lbounds_section->data_offset);
 609         greloc(cur_text_section, sym_data,
 610                ind + 1, R_386_32);
 611         oad(0xb8, 0); /* mov %eax, xxx */
 612         gen_static_call(TOK___bound_local_new);
 613
 614         ind = saved_ind;
 615         /* generate bound check local freeing */
 616         o(0x5250); /* save returned value, if any */
 617         greloc(cur_text_section, sym_data,
 618                ind + 1, R_386_32);
 619         oad(0xb8, 0); /* mov %eax, xxx */
 620         gen_static_call(TOK___bound_local_delete);
 621
 622         o(0x585a); /* restore returned value, if any */
 623     }
 624 #endif
 625     o(0xc9); /* leave */
 626     if (func_ret_sub == 0) {
 627         o(0xc3); /* ret */
 628     } else {
 629         o(0xc2); /* ret n */
 630         g(func_ret_sub);
 631         g(func_ret_sub >> 8);
 632     }
 633     /* align local size to word & save local variables */
 634
 635     v = (-loc + 3) & -4;
 636     saved_ind = ind;
 637     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 638 #ifdef TCC_TARGET_PE
 639     if (v >= 4096) {
 640         oad(0xb8, v); /* mov stacksize, %eax */
 641         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 642     } else
 643 #endif
 644     {
 645         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 646         o(0xec81);  /* sub esp, stacksize */
 647         gen_le32(v);
 648 #if FUNC_PROLOG_SIZE == 10
 649         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 650 #endif
 651     }
 652     ind = saved_ind;
 653 }
 654
 655 /* generate a jump to a label */
 656 ST_FUNC int gjmp(int t)
 657 {
 658     return psym(0xe9, t);
 659 }
 660
 661 /* generate a jump to a fixed address */
 662 ST_FUNC void gjmp_addr(int a)
 663 {
 664     int r;
 665     r = a - ind - 2;
 666     if (r == (char)r) {
 667         g(0xeb);
 668         g(r);
 669     } else {
 670         oad(0xe9, a - ind - 5);
 671     }
 672 }
 673
 674 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 675 ST_FUNC int gtst(int inv, int t)
 676 {
 677     int v, *p;
 678
 679     v = vtop->r & VT_VALMASK;
 680     if (v == VT_CMP) {
 681         /* fast case : can jump directly since flags are set */
 682         g(0x0f);
 683         t = psym((vtop->c.i - 16) ^ inv, t);
 684     } else if (v == VT_JMP || v == VT_JMPI) {
 685         /* && or || optimization */
 686         if ((v & 1) == inv) {
 687             /* insert vtop->c jump list in t */
 688             p = &vtop->c.i;
 689             while (*p != 0)
 690                 p = (int *)(cur_text_section->data + *p);
 691             *p = t;
 692             t = vtop->c.i;
 693         } else {
 694             t = gjmp(t);
 695             gsym(vtop->c.i);
 696         }
 697     } else {
 698         if (is_float(vtop->type.t) ||
 699             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 700             vpushi(0);
 701             gen_op(TOK_NE);
 702         }
 703         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 704             /* constant jmp optimization */
 705             if ((vtop->c.i != 0) != inv)
 706                 t = gjmp(t);
 707         } else {
 708             v = gv(RC_INT);
 709             o(0x85);
 710             o(0xc0 + v * 9);
 711             g(0x0f);
 712             t = psym(0x85 ^ inv, t);
 713         }
 714     }
 715     vtop--;
 716     return t;
 717 }
 718
 719 /* generate an integer binary operation */
 720 ST_FUNC void gen_opi(int op)
 721 {
 722     int r, fr, opc, c;
 723
 724     switch(op) {
 725     case '+':
 726     case TOK_ADDC1: /* add with carry generation */
 727         opc = 0;
 728     gen_op8:
 729         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 730             /* constant case */
 731             vswap();
 732             r = gv(RC_INT);
 733             vswap();
 734             c = vtop->c.i;
 735             if (c == (char)c) {
 736                 /* generate inc and dec for smaller code */
 737                 if (c==1 && opc==0) {
 738                     o (0x40 | r); // inc
 739                 } else if (c==1 && opc==5) {
 740                     o (0x48 | r); // dec
 741                 } else {
 742                     o(0x83);
 743                     o(0xc0 | (opc << 3) | r);
 744                     g(c);
 745                 }
 746             } else {
 747                 o(0x81);
 748                 oad(0xc0 | (opc << 3) | r, c);
 749             }
 750         } else {
 751             gv2(RC_INT, RC_INT);
 752             r = vtop[-1].r;
 753             fr = vtop[0].r;
 754             o((opc << 3) | 0x01);
 755             o(0xc0 + r + fr * 8);
 756         }
 757         vtop--;
 758         if (op >= TOK_ULT && op <= TOK_GT) {
 759             vtop->r = VT_CMP;
 760             vtop->c.i = op;
 761         }
 762         break;
 763     case '-':
 764     case TOK_SUBC1: /* sub with carry generation */
 765         opc = 5;
 766         goto gen_op8;
 767     case TOK_ADDC2: /* add with carry use */
 768         opc = 2;
 769         goto gen_op8;
 770     case TOK_SUBC2: /* sub with carry use */
 771         opc = 3;
 772         goto gen_op8;
 773     case '&':
 774         opc = 4;
 775         goto gen_op8;
 776     case '^':
 777         opc = 6;
 778         goto gen_op8;
 779     case '|':
 780         opc = 1;
 781         goto gen_op8;
 782     case '*':
 783         gv2(RC_INT, RC_INT);
 784         r = vtop[-1].r;
 785         fr = vtop[0].r;
 786         vtop--;
 787         o(0xaf0f); /* imul fr, r */
 788         o(0xc0 + fr + r * 8);
 789         break;
 790     case TOK_SHL:
 791         opc = 4;
 792         goto gen_shift;
 793     case TOK_SHR:
 794         opc = 5;
 795         goto gen_shift;
 796     case TOK_SAR:
 797         opc = 7;
 798     gen_shift:
 799         opc = 0xc0 | (opc << 3);
 800         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 801             /* constant case */
 802             vswap();
 803             r = gv(RC_INT);
 804             vswap();
 805             c = vtop->c.i & 0x1f;
 806             o(0xc1); /* shl/shr/sar $xxx, r */
 807             o(opc | r);
 808             g(c);
 809         } else {
 810             /* we generate the shift in ecx */
 811             gv2(RC_INT, RC_ECX);
 812             r = vtop[-1].r;
 813             o(0xd3); /* shl/shr/sar %cl, r */
 814             o(opc | r);
 815         }
 816         vtop--;
 817         break;
 818     case '/':
 819     case TOK_UDIV:
 820     case TOK_PDIV:
 821     case '%':
 822     case TOK_UMOD:
 823     case TOK_UMULL:
 824         /* first operand must be in eax */
 825         /* XXX: need better constraint for second operand */
 826         gv2(RC_EAX, RC_ECX);
 827         r = vtop[-1].r;
 828         fr = vtop[0].r;
 829         vtop--;
 830         save_reg(TREG_EDX);
 831         if (op == TOK_UMULL) {
 832             o(0xf7); /* mul fr */
 833             o(0xe0 + fr);
 834             vtop->r2 = TREG_EDX;
 835             r = TREG_EAX;
 836         } else {
 837             if (op == TOK_UDIV || op == TOK_UMOD) {
 838                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 839                 o(0xf0 + fr);
 840             } else {
 841                 o(0xf799); /* cltd, idiv fr, %eax */
 842                 o(0xf8 + fr);
 843             }
 844             if (op == '%' || op == TOK_UMOD)
 845                 r = TREG_EDX;
 846             else
 847                 r = TREG_EAX;
 848         }
 849         vtop->r = r;
 850         break;
 851     default:
 852         opc = 7;
 853         goto gen_op8;
 854     }
 855 }
 856
 857 /* generate a floating point operation 'v = t1 op t2' instruction. The
 858    two operands are guaranted to have the same floating point type */
 859 /* XXX: need to use ST1 too */
 860 ST_FUNC void gen_opf(int op)
 861 {
 862     int a, ft, fc, swapped, r;
 863
 864     /* convert constants to memory references */
 865     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 866         vswap();
 867         gv(RC_FLOAT);
 868         vswap();
 869     }
 870     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 871         gv(RC_FLOAT);
 872
 873     /* must put at least one value in the floating point register */
 874     if ((vtop[-1].r & VT_LVAL) &&
 875         (vtop[0].r & VT_LVAL)) {
 876         vswap();
 877         gv(RC_FLOAT);
 878         vswap();
 879     }
 880     swapped = 0;
 881     /* swap the stack if needed so that t1 is the register and t2 is
 882        the memory reference */
 883     if (vtop[-1].r & VT_LVAL) {
 884         vswap();
 885         swapped = 1;
 886     }
 887     if (op >= TOK_ULT && op <= TOK_GT) {
 888         /* load on stack second operand */
 889         load(TREG_ST0, vtop);
 890         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 891         if (op == TOK_GE || op == TOK_GT)
 892             swapped = !swapped;
 893         else if (op == TOK_EQ || op == TOK_NE)
 894             swapped = 0;
 895         if (swapped)
 896             o(0xc9d9); /* fxch %st(1) */
 897         o(0xe9da); /* fucompp */
 898         o(0xe0df); /* fnstsw %ax */
 899         if (op == TOK_EQ) {
 900             o(0x45e480); /* and $0x45, %ah */
 901             o(0x40fC80); /* cmp $0x40, %ah */
 902         } else if (op == TOK_NE) {
 903             o(0x45e480); /* and $0x45, %ah */
 904             o(0x40f480); /* xor $0x40, %ah */
 905             op = TOK_NE;
 906         } else if (op == TOK_GE || op == TOK_LE) {
 907             o(0x05c4f6); /* test $0x05, %ah */
 908             op = TOK_EQ;
 909         } else {
 910             o(0x45c4f6); /* test $0x45, %ah */
 911             op = TOK_EQ;
 912         }
 913         vtop--;
 914         vtop->r = VT_CMP;
 915         vtop->c.i = op;
 916     } else {
 917         /* no memory reference possible for long double operations */
 918         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 919             load(TREG_ST0, vtop);
 920             swapped = !swapped;
 921         }
 922
 923         switch(op) {
 924         default:
 925         case '+':
 926             a = 0;
 927             break;
 928         case '-':
 929             a = 4;
 930             if (swapped)
 931                 a++;
 932             break;
 933         case '*':
 934             a = 1;
 935             break;
 936         case '/':
 937             a = 6;
 938             if (swapped)
 939                 a++;
 940             break;
 941         }
 942         ft = vtop->type.t;
 943         fc = vtop->c.ul;
 944         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 945             o(0xde); /* fxxxp %st, %st(1) */
 946             o(0xc1 + (a << 3));
 947         } else {
 948             /* if saved lvalue, then we must reload it */
 949             r = vtop->r;
 950             if ((r & VT_VALMASK) == VT_LLOCAL) {
 951                 SValue v1;
 952                 r = get_reg(RC_INT);
 953                 v1.type.t = VT_INT;
 954                 v1.r = VT_LOCAL | VT_LVAL;
 955                 v1.c.ul = fc;
 956                 load(r, &v1);
 957                 fc = 0;
 958             }
 959
 960             if ((ft & VT_BTYPE) == VT_DOUBLE)
 961                 o(0xdc);
 962             else
 963                 o(0xd8);
 964             gen_modrm(a, r, vtop->sym, fc);
 965         }
 966         vtop--;
 967     }
 968 }
 969
 970 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 971    and 'long long' cases. */
 972 ST_FUNC void gen_cvt_itof(int t)
 973 {
 974     save_reg(TREG_ST0);
 975     gv(RC_INT);
 976     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 977         /* signed long long to float/double/long double (unsigned case
 978            is handled generically) */
 979         o(0x50 + vtop->r2); /* push r2 */
 980         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 981         o(0x242cdf); /* fildll (%esp) */
 982         o(0x08c483); /* add $8, %esp */
 983     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 984                (VT_INT | VT_UNSIGNED)) {
 985         /* unsigned int to float/double/long double */
 986         o(0x6a); /* push $0 */
 987         g(0x00);
 988         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 989         o(0x242cdf); /* fildll (%esp) */
 990         o(0x08c483); /* add $8, %esp */
 991     } else {
 992         /* int to float/double/long double */
 993         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 994         o(0x2404db); /* fildl (%esp) */
 995         o(0x04c483); /* add $4, %esp */
 996     }
 997     vtop->r = TREG_ST0;
 998 }
 999
1000 /* convert fp to int 't' type */
1001 /* XXX: handle long long case */
1002 ST_FUNC void gen_cvt_ftoi(int t)
1003 {
1004     gv(RC_FLOAT);
1005     save_reg(TREG_EAX);
1006     save_reg(TREG_EDX);
1007     gen_static_call(TOK___tcc_cvt_ftol);
1008     vtop->r = TREG_EAX; /* mark reg as used */
1009     if (t == VT_LLONG)
1010         vtop->r2 = TREG_EDX;
1011 }
1012
1013 /* convert from one floating point type to another */
1014 ST_FUNC void gen_cvt_ftof(int t)
1015 {
1016     /* all we have to do on i386 is to put the float in a register */
1017     gv(RC_FLOAT);
1018 }
1019
1020 /* computed goto support */
1021 ST_FUNC void ggoto(void)
1022 {
1023     gcall_or_jmp(1);
1024     vtop--;
1025 }
1026
1027 /* bound check support functions */
1028 #ifdef CONFIG_TCC_BCHECK
1029
1030 /* generate a bounded pointer addition */
1031 ST_FUNC void gen_bounded_ptr_add(void)
1032 {
1033     /* prepare fast i386 function call (args in eax and edx) */
1034     gv2(RC_EAX, RC_EDX);
1035     /* save all temporary registers */
1036     vtop -= 2;
1037     save_regs(0);
1038     /* do a fast function call */
1039     gen_static_call(TOK___bound_ptr_add);
1040     /* returned pointer is in eax */
1041     vtop++;
1042     vtop->r = TREG_EAX | VT_BOUNDED;
1043     /* address of bounding function call point */
1044     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1045 }
1046
1047 /* patch pointer addition in vtop so that pointer dereferencing is
1048    also tested */
1049 ST_FUNC void gen_bounded_ptr_deref(void)
1050 {
1051     int func;
1052     int size, align;
1053     Elf32_Rel *rel;
1054     Sym *sym;
1055
1056     size = 0;
1057     /* XXX: put that code in generic part of tcc */
1058     if (!is_float(vtop->type.t)) {
1059         if (vtop->r & VT_LVAL_BYTE)
1060             size = 1;
1061         else if (vtop->r & VT_LVAL_SHORT)
1062             size = 2;
1063     }
1064     if (!size)
1065         size = type_size(&vtop->type, &align);
1066     switch(size) {
1067     case  1: func = TOK___bound_ptr_indir1; break;
1068     case  2: func = TOK___bound_ptr_indir2; break;
1069     case  4: func = TOK___bound_ptr_indir4; break;
1070     case  8: func = TOK___bound_ptr_indir8; break;
1071     case 12: func = TOK___bound_ptr_indir12; break;
1072     case 16: func = TOK___bound_ptr_indir16; break;
1073     default:
1074         tcc_error("unhandled size when dereferencing bounded pointer");
1075         func = 0;
1076         break;
1077     }
1078
1079     /* patch relocation */
1080     /* XXX: find a better solution ? */
1081     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1082     sym = external_global_sym(func, &func_old_type, 0);
1083     if (!sym->c)
1084         put_extern_sym(sym, NULL, 0, 0);
1085     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1086 }
1087 #endif
1088
1089 /* Save the stack pointer onto the stack */
1090 ST_FUNC void gen_vla_sp_save(int addr) {
1091     /* mov %esp,addr(%ebp)*/
1092     o(0x89);
1093     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1094 }
1095
1096 /* Restore the SP from a location on the stack */
1097 ST_FUNC void gen_vla_sp_restore(int addr) {
1098     o(0x8b);
1099     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1100 }
1101
1102 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1103 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1104 #ifdef TCC_TARGET_PE
1105     /* alloca does more than just adjust %rsp on Windows */
1106     vpush_global_sym(&func_old_type, TOK_alloca);
1107     vswap(); /* Move alloca ref past allocation size */
1108     gfunc_call(1);
1109     vset(type, REG_IRET, 0);
1110 #else
1111     int r;
1112     r = gv(RC_INT); /* allocation size */
1113     /* sub r,%rsp */
1114     o(0x2b);
1115     o(0xe0 | r);
1116     /* We align to 16 bytes rather than align */
1117     /* and ~15, %esp */
1118     o(0xf0e483);
1119     /* mov %esp, r */
1120     o(0x89);
1121     o(0xe0 | r);
1122     vpop();
1123     vset(type, r, 0);
1124 #endif
1125 }
1126
1127 /* end of X86 code generator */
1128 /*************************************************************/
1129 #endif
1130 /*************************************************************/