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 };
  47
  48 /* return registers for function */
  49 #define REG_IRET TREG_EAX /* single word int return register */
  50 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  51 #define REG_FRET TREG_ST0 /* float return register */
  52
  53 /* defined if function parameters must be evaluated in reverse order */
  54 #define INVERT_FUNC_PARAMS
  55
  56 /* defined if structures are passed as pointers. Otherwise structures
  57    are directly pushed on stack. */
  58 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  59
  60 /* pointer size, in bytes */
  61 #define PTR_SIZE 4
  62
  63 /* long double size and alignment, in bytes */
  64 #define LDOUBLE_SIZE  12
  65 #define LDOUBLE_ALIGN 4
  66 /* maximum alignment (for aligned attribute support) */
  67 #define MAX_ALIGN     8
  68
  69
  70 #define psym oad
  71
  72 /******************************************************/
  73 /* ELF defines */
  74
  75 #define EM_TCC_TARGET EM_386
  76
  77 /* relocation type for 32 bit data relocation */
  78 #define R_DATA_32   R_386_32
  79 #define R_DATA_PTR  R_386_32
  80 #define R_JMP_SLOT  R_386_JMP_SLOT
  81 #define R_COPY      R_386_COPY
  82
  83 #define ELF_START_ADDR 0x08048000
  84 #define ELF_PAGE_SIZE  0x1000
  85
  86 /******************************************************/
  87 #else /* ! TARGET_DEFS_ONLY */
  88 /******************************************************/
  89 #include "tcc.h"
  90
  91 ST_DATA const int reg_classes[NB_REGS] = {
  92     /* eax */ RC_INT | RC_EAX,
  93     /* ecx */ RC_INT | RC_ECX,
  94     /* edx */ RC_INT | RC_EDX,
  95     /* st0 */ RC_FLOAT | RC_ST0,
  96 };
  97
  98 static unsigned long func_sub_sp_offset;
  99 static int func_ret_sub;
 100 #ifdef CONFIG_TCC_BCHECK
 101 static unsigned long func_bound_offset;
 102 #endif
 103
 104 /* XXX: make it faster ? */
 105 ST_FUNC void g(int c)
 106 {
 107     int ind1;
 108     ind1 = ind + 1;
 109     if (ind1 > cur_text_section->data_allocated)
 110         section_realloc(cur_text_section, ind1);
 111     cur_text_section->data[ind] = c;
 112     ind = ind1;
 113 }
 114
 115 ST_FUNC void o(unsigned int c)
 116 {
 117     while (c) {
 118         g(c);
 119         c = c >> 8;
 120     }
 121 }
 122
 123 ST_FUNC void gen_le16(int v)
 124 {
 125     g(v);
 126     g(v >> 8);
 127 }
 128
 129 ST_FUNC void gen_le32(int c)
 130 {
 131     g(c);
 132     g(c >> 8);
 133     g(c >> 16);
 134     g(c >> 24);
 135 }
 136
 137 /* output a symbol and patch all calls to it */
 138 ST_FUNC void gsym_addr(int t, int a)
 139 {
 140     int n, *ptr;
 141     while (t) {
 142         ptr = (int *)(cur_text_section->data + t);
 143         n = *ptr; /* next value */
 144         *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     *(int *)(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.ul;
 225
 226     v = fr & VT_VALMASK;
 227     if (fr & VT_LVAL) {
 228         if (v == VT_LLOCAL) {
 229             v1.type.t = VT_INT;
 230             v1.r = VT_LOCAL | VT_LVAL;
 231             v1.c.ul = fc;
 232             fr = r;
 233             if (!(reg_classes[fr] & RC_INT))
 234                 fr = get_reg(RC_INT);
 235             load(fr, &v1);
 236         }
 237         if ((ft & VT_BTYPE) == VT_FLOAT) {
 238             o(0xd9); /* flds */
 239             r = 0;
 240         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 241             o(0xdd); /* fldl */
 242             r = 0;
 243         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 244             o(0xdb); /* fldt */
 245             r = 5;
 246         } else if ((ft & VT_TYPE) == VT_BYTE) {
 247             o(0xbe0f);   /* movsbl */
 248         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 249             o(0xb60f);   /* movzbl */
 250         } else if ((ft & VT_TYPE) == VT_SHORT) {
 251             o(0xbf0f);   /* movswl */
 252         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 253             o(0xb70f);   /* movzwl */
 254         } else {
 255             o(0x8b);     /* movl */
 256         }
 257         gen_modrm(r, fr, sv->sym, fc);
 258     } else {
 259         if (v == VT_CONST) {
 260             o(0xb8 + r); /* mov $xx, r */
 261             gen_addr32(fr, sv->sym, fc);
 262         } else if (v == VT_LOCAL) {
 263             if (fc) {
 264                 o(0x8d); /* lea xxx(%ebp), r */
 265                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 266             } else {
 267                 o(0x89);
 268                 o(0xe8 + r); /* mov %ebp, r */
 269             }
 270         } else if (v == VT_CMP) {
 271             oad(0xb8 + r, 0); /* mov $0, r */
 272             o(0x0f); /* setxx %br */
 273             o(fc);
 274             o(0xc0 + r);
 275         } else if (v == VT_JMP || v == VT_JMPI) {
 276             t = v & 1;
 277             oad(0xb8 + r, t); /* mov $1, r */
 278             o(0x05eb); /* jmp after */
 279             gsym(fc);
 280             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 281         } else if (v != r) {
 282             o(0x89);
 283             o(0xc0 + r + v * 8); /* mov v, r */
 284         }
 285     }
 286 }
 287
 288 /* store register 'r' in lvalue 'v' */
 289 ST_FUNC void store(int r, SValue *v)
 290 {
 291     int fr, bt, ft, fc;
 292
 293 #ifdef TCC_TARGET_PE
 294     SValue v2;
 295     v = pe_getimport(v, &v2);
 296 #endif
 297
 298     ft = v->type.t;
 299     fc = v->c.ul;
 300     fr = v->r & VT_VALMASK;
 301     bt = ft & VT_BTYPE;
 302     /* XXX: incorrect if float reg to reg */
 303     if (bt == VT_FLOAT) {
 304         o(0xd9); /* fsts */
 305         r = 2;
 306     } else if (bt == VT_DOUBLE) {
 307         o(0xdd); /* fstpl */
 308         r = 2;
 309     } else if (bt == VT_LDOUBLE) {
 310         o(0xc0d9); /* fld %st(0) */
 311         o(0xdb); /* fstpt */
 312         r = 7;
 313     } else {
 314         if (bt == VT_SHORT)
 315             o(0x66);
 316         if (bt == VT_BYTE || bt == VT_BOOL)
 317             o(0x88);
 318         else
 319             o(0x89);
 320     }
 321     if (fr == VT_CONST ||
 322         fr == VT_LOCAL ||
 323         (v->r & VT_LVAL)) {
 324         gen_modrm(r, v->r, v->sym, fc);
 325     } else if (fr != r) {
 326         o(0xc0 + fr + r * 8); /* mov r, fr */
 327     }
 328 }
 329
 330 static void gadd_sp(int val)
 331 {
 332     if (val == (char)val) {
 333         o(0xc483);
 334         g(val);
 335     } else {
 336         oad(0xc481, val); /* add $xxx, %esp */
 337     }
 338 }
 339
 340 /* 'is_jmp' is '1' if it is a jump */
 341 static void gcall_or_jmp(int is_jmp)
 342 {
 343     int r;
 344     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 345         /* constant case */
 346         if (vtop->r & VT_SYM) {
 347             /* relocation case */
 348             greloc(cur_text_section, vtop->sym,
 349                    ind + 1, R_386_PC32);
 350         } else {
 351             /* put an empty PC32 relocation */
 352             put_elf_reloc(symtab_section, cur_text_section,
 353                           ind + 1, R_386_PC32, 0);
 354         }
 355         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 356     } else {
 357         /* otherwise, indirect call */
 358         r = gv(RC_INT);
 359         o(0xff); /* call/jmp *r */
 360         o(0xd0 + r + (is_jmp << 4));
 361     }
 362 }
 363
 364 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 365 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 366
 367 /* Return 1 if this function returns via an sret pointer, 0 otherwise */
 368 ST_FUNC int gfunc_sret(CType *vt, CType *ret, int *ret_align) {
 369     *ret_align = 1; // Never have to re-align return values for x86
 370 #ifdef TCC_TARGET_PE
 371     int size, align;
 372     size = type_size(vt, &align);
 373     if (size > 8) {
 374         return 1;
 375     } else if (size > 4) {
 376         ret->ref = NULL;
 377         ret->t = VT_LLONG;
 378         return 0;
 379     } else {
 380         ret->ref = NULL;
 381         ret->t = VT_INT;
 382         return 0;
 383     }
 384 #else
 385     return 1;
 386 #endif
 387 }
 388
 389 /* Generate function call. The function address is pushed first, then
 390    all the parameters in call order. This functions pops all the
 391    parameters and the function address. */
 392 ST_FUNC void gfunc_call(int nb_args)
 393 {
 394     int size, align, r, args_size, i, func_call;
 395     Sym *func_sym;
 396
 397     args_size = 0;
 398     for(i = 0;i < nb_args; i++) {
 399         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 400             size = type_size(&vtop->type, &align);
 401             /* align to stack align size */
 402             size = (size + 3) & ~3;
 403             /* allocate the necessary size on stack */
 404             oad(0xec81, size); /* sub $xxx, %esp */
 405             /* generate structure store */
 406             r = get_reg(RC_INT);
 407             o(0x89); /* mov %esp, r */
 408             o(0xe0 + r);
 409             vset(&vtop->type, r | VT_LVAL, 0);
 410             vswap();
 411             vstore();
 412             args_size += size;
 413         } else if (is_float(vtop->type.t)) {
 414             gv(RC_FLOAT); /* only one float register */
 415             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 416                 size = 4;
 417             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 418                 size = 8;
 419             else
 420                 size = 12;
 421             oad(0xec81, size); /* sub $xxx, %esp */
 422             if (size == 12)
 423                 o(0x7cdb);
 424             else
 425                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 426             g(0x24);
 427             g(0x00);
 428             args_size += size;
 429         } else {
 430             /* simple type (currently always same size) */
 431             /* XXX: implicit cast ? */
 432             r = gv(RC_INT);
 433             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 434                 size = 8;
 435                 o(0x50 + vtop->r2); /* push r */
 436             } else {
 437                 size = 4;
 438             }
 439             o(0x50 + r); /* push r */
 440             args_size += size;
 441         }
 442         vtop--;
 443     }
 444     save_regs(0); /* save used temporary registers */
 445     func_sym = vtop->type.ref;
 446     func_call = FUNC_CALL(func_sym->r);
 447     /* fast call case */
 448     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 449         func_call == FUNC_FASTCALLW) {
 450         int fastcall_nb_regs;
 451         uint8_t *fastcall_regs_ptr;
 452         if (func_call == FUNC_FASTCALLW) {
 453             fastcall_regs_ptr = fastcallw_regs;
 454             fastcall_nb_regs = 2;
 455         } else {
 456             fastcall_regs_ptr = fastcall_regs;
 457             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 458         }
 459         for(i = 0;i < fastcall_nb_regs; i++) {
 460             if (args_size <= 0)
 461                 break;
 462             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 463             /* XXX: incorrect for struct/floats */
 464             args_size -= 4;
 465         }
 466     }
 467 #ifndef TCC_TARGET_PE
 468     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 469         args_size -= 4;
 470 #endif
 471     gcall_or_jmp(0);
 472
 473     if (args_size && func_call != FUNC_STDCALL)
 474         gadd_sp(args_size);
 475     vtop--;
 476 }
 477
 478 #ifdef TCC_TARGET_PE
 479 #define FUNC_PROLOG_SIZE 10
 480 #else
 481 #define FUNC_PROLOG_SIZE 9
 482 #endif
 483
 484 /* generate function prolog of type 't' */
 485 ST_FUNC void gfunc_prolog(CType *func_type)
 486 {
 487     int addr, align, size, func_call, fastcall_nb_regs;
 488     int param_index, param_addr;
 489     uint8_t *fastcall_regs_ptr;
 490     Sym *sym;
 491     CType *type;
 492
 493     sym = func_type->ref;
 494     func_call = FUNC_CALL(sym->r);
 495     addr = 8;
 496     loc = 0;
 497     func_vc = 0;
 498
 499     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 500         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 501         fastcall_regs_ptr = fastcall_regs;
 502     } else if (func_call == FUNC_FASTCALLW) {
 503         fastcall_nb_regs = 2;
 504         fastcall_regs_ptr = fastcallw_regs;
 505     } else {
 506         fastcall_nb_regs = 0;
 507         fastcall_regs_ptr = NULL;
 508     }
 509     param_index = 0;
 510
 511     ind += FUNC_PROLOG_SIZE;
 512     func_sub_sp_offset = ind;
 513     /* if the function returns a structure, then add an
 514        implicit pointer parameter */
 515     func_vt = sym->type;
 516 #ifdef TCC_TARGET_PE
 517     size = type_size(&func_vt,&align);
 518     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 519 #else
 520     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 521 #endif
 522         /* XXX: fastcall case ? */
 523         func_vc = addr;
 524         addr += 4;
 525         param_index++;
 526     }
 527     /* define parameters */
 528     while ((sym = sym->next) != NULL) {
 529         type = &sym->type;
 530         size = type_size(type, &align);
 531         size = (size + 3) & ~3;
 532 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 533         /* structs are passed as pointer */
 534         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 535             size = 4;
 536         }
 537 #endif
 538         if (param_index < fastcall_nb_regs) {
 539             /* save FASTCALL register */
 540             loc -= 4;
 541             o(0x89);     /* movl */
 542             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 543             param_addr = loc;
 544         } else {
 545             param_addr = addr;
 546             addr += size;
 547         }
 548         sym_push(sym->v & ~SYM_FIELD, type,
 549                  VT_LOCAL | lvalue_type(type->t), param_addr);
 550         param_index++;
 551     }
 552     func_ret_sub = 0;
 553     /* pascal type call ? */
 554     if (func_call == FUNC_STDCALL)
 555         func_ret_sub = addr - 8;
 556 #ifndef TCC_TARGET_PE
 557     else if (func_vc)
 558         func_ret_sub = 4;
 559 #endif
 560
 561 #ifdef CONFIG_TCC_BCHECK
 562     /* leave some room for bound checking code */
 563     if (tcc_state->do_bounds_check) {
 564         oad(0xb8, 0); /* lbound section pointer */
 565         oad(0xb8, 0); /* call to function */
 566         func_bound_offset = lbounds_section->data_offset;
 567     }
 568 #endif
 569 }
 570
 571 /* generate function epilog */
 572 ST_FUNC void gfunc_epilog(void)
 573 {
 574     int v, saved_ind;
 575
 576 #ifdef CONFIG_TCC_BCHECK
 577     if (tcc_state->do_bounds_check
 578      && func_bound_offset != lbounds_section->data_offset) {
 579         int saved_ind;
 580         int *bounds_ptr;
 581         Sym *sym, *sym_data;
 582         /* add end of table info */
 583         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 584         *bounds_ptr = 0;
 585         /* generate bound local allocation */
 586         saved_ind = ind;
 587         ind = func_sub_sp_offset;
 588         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 589                                func_bound_offset, lbounds_section->data_offset);
 590         greloc(cur_text_section, sym_data,
 591                ind + 1, R_386_32);
 592         oad(0xb8, 0); /* mov %eax, xxx */
 593         sym = external_global_sym(TOK___bound_local_new, &func_old_type, 0);
 594         greloc(cur_text_section, sym,
 595                ind + 1, R_386_PC32);
 596         oad(0xe8, -4);
 597         ind = saved_ind;
 598         /* generate bound check local freeing */
 599         o(0x5250); /* save returned value, if any */
 600         greloc(cur_text_section, sym_data,
 601                ind + 1, R_386_32);
 602         oad(0xb8, 0); /* mov %eax, xxx */
 603         sym = external_global_sym(TOK___bound_local_delete, &func_old_type, 0);
 604         greloc(cur_text_section, sym,
 605                ind + 1, R_386_PC32);
 606         oad(0xe8, -4);
 607         o(0x585a); /* restore returned value, if any */
 608     }
 609 #endif
 610     o(0xc9); /* leave */
 611     if (func_ret_sub == 0) {
 612         o(0xc3); /* ret */
 613     } else {
 614         o(0xc2); /* ret n */
 615         g(func_ret_sub);
 616         g(func_ret_sub >> 8);
 617     }
 618     /* align local size to word & save local variables */
 619
 620     v = (-loc + 3) & -4;
 621     saved_ind = ind;
 622     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 623 #ifdef TCC_TARGET_PE
 624     if (v >= 4096) {
 625         Sym *sym = external_global_sym(TOK___chkstk, &func_old_type, 0);
 626         oad(0xb8, v); /* mov stacksize, %eax */
 627         oad(0xe8, -4); /* call __chkstk, (does the stackframe too) */
 628         greloc(cur_text_section, sym, ind-4, R_386_PC32);
 629     } else
 630 #endif
 631     {
 632         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 633         o(0xec81);  /* sub esp, stacksize */
 634         gen_le32(v);
 635 #if FUNC_PROLOG_SIZE == 10
 636         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 637 #endif
 638     }
 639     ind = saved_ind;
 640 }
 641
 642 /* generate a jump to a label */
 643 ST_FUNC int gjmp(int t)
 644 {
 645     return psym(0xe9, t);
 646 }
 647
 648 /* generate a jump to a fixed address */
 649 ST_FUNC void gjmp_addr(int a)
 650 {
 651     int r;
 652     r = a - ind - 2;
 653     if (r == (char)r) {
 654         g(0xeb);
 655         g(r);
 656     } else {
 657         oad(0xe9, a - ind - 5);
 658     }
 659 }
 660
 661 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 662 ST_FUNC int gtst(int inv, int t)
 663 {
 664     int v, *p;
 665
 666     v = vtop->r & VT_VALMASK;
 667     if (v == VT_CMP) {
 668         /* fast case : can jump directly since flags are set */
 669         g(0x0f);
 670         t = psym((vtop->c.i - 16) ^ inv, t);
 671     } else if (v == VT_JMP || v == VT_JMPI) {
 672         /* && or || optimization */
 673         if ((v & 1) == inv) {
 674             /* insert vtop->c jump list in t */
 675             p = &vtop->c.i;
 676             while (*p != 0)
 677                 p = (int *)(cur_text_section->data + *p);
 678             *p = t;
 679             t = vtop->c.i;
 680         } else {
 681             t = gjmp(t);
 682             gsym(vtop->c.i);
 683         }
 684     } else {
 685         if (is_float(vtop->type.t) ||
 686             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 687             vpushi(0);
 688             gen_op(TOK_NE);
 689         }
 690         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 691             /* constant jmp optimization */
 692             if ((vtop->c.i != 0) != inv)
 693                 t = gjmp(t);
 694         } else {
 695             v = gv(RC_INT);
 696             o(0x85);
 697             o(0xc0 + v * 9);
 698             g(0x0f);
 699             t = psym(0x85 ^ inv, t);
 700         }
 701     }
 702     vtop--;
 703     return t;
 704 }
 705
 706 /* generate an integer binary operation */
 707 ST_FUNC void gen_opi(int op)
 708 {
 709     int r, fr, opc, c;
 710
 711     switch(op) {
 712     case '+':
 713     case TOK_ADDC1: /* add with carry generation */
 714         opc = 0;
 715     gen_op8:
 716         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 717             /* constant case */
 718             vswap();
 719             r = gv(RC_INT);
 720             vswap();
 721             c = vtop->c.i;
 722             if (c == (char)c) {
 723                 /* generate inc and dec for smaller code */
 724                 if (c==1 && opc==0) {
 725                     o (0x40 | r); // inc
 726                 } else if (c==1 && opc==5) {
 727                     o (0x48 | r); // dec
 728                 } else {
 729                     o(0x83);
 730                     o(0xc0 | (opc << 3) | r);
 731                     g(c);
 732                 }
 733             } else {
 734                 o(0x81);
 735                 oad(0xc0 | (opc << 3) | r, c);
 736             }
 737         } else {
 738             gv2(RC_INT, RC_INT);
 739             r = vtop[-1].r;
 740             fr = vtop[0].r;
 741             o((opc << 3) | 0x01);
 742             o(0xc0 + r + fr * 8);
 743         }
 744         vtop--;
 745         if (op >= TOK_ULT && op <= TOK_GT) {
 746             vtop->r = VT_CMP;
 747             vtop->c.i = op;
 748         }
 749         break;
 750     case '-':
 751     case TOK_SUBC1: /* sub with carry generation */
 752         opc = 5;
 753         goto gen_op8;
 754     case TOK_ADDC2: /* add with carry use */
 755         opc = 2;
 756         goto gen_op8;
 757     case TOK_SUBC2: /* sub with carry use */
 758         opc = 3;
 759         goto gen_op8;
 760     case '&':
 761         opc = 4;
 762         goto gen_op8;
 763     case '^':
 764         opc = 6;
 765         goto gen_op8;
 766     case '|':
 767         opc = 1;
 768         goto gen_op8;
 769     case '*':
 770         gv2(RC_INT, RC_INT);
 771         r = vtop[-1].r;
 772         fr = vtop[0].r;
 773         vtop--;
 774         o(0xaf0f); /* imul fr, r */
 775         o(0xc0 + fr + r * 8);
 776         break;
 777     case TOK_SHL:
 778         opc = 4;
 779         goto gen_shift;
 780     case TOK_SHR:
 781         opc = 5;
 782         goto gen_shift;
 783     case TOK_SAR:
 784         opc = 7;
 785     gen_shift:
 786         opc = 0xc0 | (opc << 3);
 787         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 788             /* constant case */
 789             vswap();
 790             r = gv(RC_INT);
 791             vswap();
 792             c = vtop->c.i & 0x1f;
 793             o(0xc1); /* shl/shr/sar $xxx, r */
 794             o(opc | r);
 795             g(c);
 796         } else {
 797             /* we generate the shift in ecx */
 798             gv2(RC_INT, RC_ECX);
 799             r = vtop[-1].r;
 800             o(0xd3); /* shl/shr/sar %cl, r */
 801             o(opc | r);
 802         }
 803         vtop--;
 804         break;
 805     case '/':
 806     case TOK_UDIV:
 807     case TOK_PDIV:
 808     case '%':
 809     case TOK_UMOD:
 810     case TOK_UMULL:
 811         /* first operand must be in eax */
 812         /* XXX: need better constraint for second operand */
 813         gv2(RC_EAX, RC_ECX);
 814         r = vtop[-1].r;
 815         fr = vtop[0].r;
 816         vtop--;
 817         save_reg(TREG_EDX);
 818         if (op == TOK_UMULL) {
 819             o(0xf7); /* mul fr */
 820             o(0xe0 + fr);
 821             vtop->r2 = TREG_EDX;
 822             r = TREG_EAX;
 823         } else {
 824             if (op == TOK_UDIV || op == TOK_UMOD) {
 825                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 826                 o(0xf0 + fr);
 827             } else {
 828                 o(0xf799); /* cltd, idiv fr, %eax */
 829                 o(0xf8 + fr);
 830             }
 831             if (op == '%' || op == TOK_UMOD)
 832                 r = TREG_EDX;
 833             else
 834                 r = TREG_EAX;
 835         }
 836         vtop->r = r;
 837         break;
 838     default:
 839         opc = 7;
 840         goto gen_op8;
 841     }
 842 }
 843
 844 /* generate a floating point operation 'v = t1 op t2' instruction. The
 845    two operands are guaranted to have the same floating point type */
 846 /* XXX: need to use ST1 too */
 847 ST_FUNC void gen_opf(int op)
 848 {
 849     int a, ft, fc, swapped, r;
 850
 851     /* convert constants to memory references */
 852     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 853         vswap();
 854         gv(RC_FLOAT);
 855         vswap();
 856     }
 857     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 858         gv(RC_FLOAT);
 859
 860     /* must put at least one value in the floating point register */
 861     if ((vtop[-1].r & VT_LVAL) &&
 862         (vtop[0].r & VT_LVAL)) {
 863         vswap();
 864         gv(RC_FLOAT);
 865         vswap();
 866     }
 867     swapped = 0;
 868     /* swap the stack if needed so that t1 is the register and t2 is
 869        the memory reference */
 870     if (vtop[-1].r & VT_LVAL) {
 871         vswap();
 872         swapped = 1;
 873     }
 874     if (op >= TOK_ULT && op <= TOK_GT) {
 875         /* load on stack second operand */
 876         load(TREG_ST0, vtop);
 877         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 878         if (op == TOK_GE || op == TOK_GT)
 879             swapped = !swapped;
 880         else if (op == TOK_EQ || op == TOK_NE)
 881             swapped = 0;
 882         if (swapped)
 883             o(0xc9d9); /* fxch %st(1) */
 884         o(0xe9da); /* fucompp */
 885         o(0xe0df); /* fnstsw %ax */
 886         if (op == TOK_EQ) {
 887             o(0x45e480); /* and $0x45, %ah */
 888             o(0x40fC80); /* cmp $0x40, %ah */
 889         } else if (op == TOK_NE) {
 890             o(0x45e480); /* and $0x45, %ah */
 891             o(0x40f480); /* xor $0x40, %ah */
 892             op = TOK_NE;
 893         } else if (op == TOK_GE || op == TOK_LE) {
 894             o(0x05c4f6); /* test $0x05, %ah */
 895             op = TOK_EQ;
 896         } else {
 897             o(0x45c4f6); /* test $0x45, %ah */
 898             op = TOK_EQ;
 899         }
 900         vtop--;
 901         vtop->r = VT_CMP;
 902         vtop->c.i = op;
 903     } else {
 904         /* no memory reference possible for long double operations */
 905         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 906             load(TREG_ST0, vtop);
 907             swapped = !swapped;
 908         }
 909
 910         switch(op) {
 911         default:
 912         case '+':
 913             a = 0;
 914             break;
 915         case '-':
 916             a = 4;
 917             if (swapped)
 918                 a++;
 919             break;
 920         case '*':
 921             a = 1;
 922             break;
 923         case '/':
 924             a = 6;
 925             if (swapped)
 926                 a++;
 927             break;
 928         }
 929         ft = vtop->type.t;
 930         fc = vtop->c.ul;
 931         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 932             o(0xde); /* fxxxp %st, %st(1) */
 933             o(0xc1 + (a << 3));
 934         } else {
 935             /* if saved lvalue, then we must reload it */
 936             r = vtop->r;
 937             if ((r & VT_VALMASK) == VT_LLOCAL) {
 938                 SValue v1;
 939                 r = get_reg(RC_INT);
 940                 v1.type.t = VT_INT;
 941                 v1.r = VT_LOCAL | VT_LVAL;
 942                 v1.c.ul = fc;
 943                 load(r, &v1);
 944                 fc = 0;
 945             }
 946
 947             if ((ft & VT_BTYPE) == VT_DOUBLE)
 948                 o(0xdc);
 949             else
 950                 o(0xd8);
 951             gen_modrm(a, r, vtop->sym, fc);
 952         }
 953         vtop--;
 954     }
 955 }
 956
 957 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 958    and 'long long' cases. */
 959 ST_FUNC void gen_cvt_itof(int t)
 960 {
 961     save_reg(TREG_ST0);
 962     gv(RC_INT);
 963     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 964         /* signed long long to float/double/long double (unsigned case
 965            is handled generically) */
 966         o(0x50 + vtop->r2); /* push r2 */
 967         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 968         o(0x242cdf); /* fildll (%esp) */
 969         o(0x08c483); /* add $8, %esp */
 970     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 971                (VT_INT | VT_UNSIGNED)) {
 972         /* unsigned int to float/double/long double */
 973         o(0x6a); /* push $0 */
 974         g(0x00);
 975         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 976         o(0x242cdf); /* fildll (%esp) */
 977         o(0x08c483); /* add $8, %esp */
 978     } else {
 979         /* int to float/double/long double */
 980         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 981         o(0x2404db); /* fildl (%esp) */
 982         o(0x04c483); /* add $4, %esp */
 983     }
 984     vtop->r = TREG_ST0;
 985 }
 986
 987 /* convert fp to int 't' type */
 988 /* XXX: handle long long case */
 989 ST_FUNC void gen_cvt_ftoi(int t)
 990 {
 991     int r, r2, size;
 992     Sym *sym;
 993     CType ushort_type;
 994
 995     ushort_type.t = VT_SHORT | VT_UNSIGNED;
 996     ushort_type.ref = 0;
 997
 998     gv(RC_FLOAT);
 999     if (t != VT_INT)
1000         size = 8;
1001     else
1002         size = 4;
1003
1004     o(0x2dd9); /* ldcw xxx */
1005     sym = external_global_sym(TOK___tcc_int_fpu_control,
1006                               &ushort_type, VT_LVAL);
1007     greloc(cur_text_section, sym,
1008            ind, R_386_32);
1009     gen_le32(0);
1010
1011     oad(0xec81, size); /* sub $xxx, %esp */
1012     if (size == 4)
1013         o(0x1cdb); /* fistpl */
1014     else
1015         o(0x3cdf); /* fistpll */
1016     o(0x24);
1017     o(0x2dd9); /* ldcw xxx */
1018     sym = external_global_sym(TOK___tcc_fpu_control,
1019                               &ushort_type, VT_LVAL);
1020     greloc(cur_text_section, sym,
1021            ind, R_386_32);
1022     gen_le32(0);
1023
1024     r = get_reg(RC_INT);
1025     o(0x58 + r); /* pop r */
1026     if (size == 8) {
1027         if (t == VT_LLONG) {
1028             vtop->r = r; /* mark reg as used */
1029             r2 = get_reg(RC_INT);
1030             o(0x58 + r2); /* pop r2 */
1031             vtop->r2 = r2;
1032         } else {
1033             o(0x04c483); /* add $4, %esp */
1034         }
1035     }
1036     vtop->r = r;
1037 }
1038
1039 /* convert from one floating point type to another */
1040 ST_FUNC void gen_cvt_ftof(int t)
1041 {
1042     /* all we have to do on i386 is to put the float in a register */
1043     gv(RC_FLOAT);
1044 }
1045
1046 /* computed goto support */
1047 ST_FUNC void ggoto(void)
1048 {
1049     gcall_or_jmp(1);
1050     vtop--;
1051 }
1052
1053 /* bound check support functions */
1054 #ifdef CONFIG_TCC_BCHECK
1055
1056 /* generate a bounded pointer addition */
1057 ST_FUNC void gen_bounded_ptr_add(void)
1058 {
1059     Sym *sym;
1060
1061     /* prepare fast i386 function call (args in eax and edx) */
1062     gv2(RC_EAX, RC_EDX);
1063     /* save all temporary registers */
1064     vtop -= 2;
1065     save_regs(0);
1066     /* do a fast function call */
1067     sym = external_global_sym(TOK___bound_ptr_add, &func_old_type, 0);
1068     greloc(cur_text_section, sym,
1069            ind + 1, R_386_PC32);
1070     oad(0xe8, -4);
1071     /* returned pointer is in eax */
1072     vtop++;
1073     vtop->r = TREG_EAX | VT_BOUNDED;
1074     /* address of bounding function call point */
1075     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1076 }
1077
1078 /* patch pointer addition in vtop so that pointer dereferencing is
1079    also tested */
1080 ST_FUNC void gen_bounded_ptr_deref(void)
1081 {
1082     int func;
1083     int size, align;
1084     Elf32_Rel *rel;
1085     Sym *sym;
1086
1087     size = 0;
1088     /* XXX: put that code in generic part of tcc */
1089     if (!is_float(vtop->type.t)) {
1090         if (vtop->r & VT_LVAL_BYTE)
1091             size = 1;
1092         else if (vtop->r & VT_LVAL_SHORT)
1093             size = 2;
1094     }
1095     if (!size)
1096         size = type_size(&vtop->type, &align);
1097     switch(size) {
1098     case  1: func = TOK___bound_ptr_indir1; break;
1099     case  2: func = TOK___bound_ptr_indir2; break;
1100     case  4: func = TOK___bound_ptr_indir4; break;
1101     case  8: func = TOK___bound_ptr_indir8; break;
1102     case 12: func = TOK___bound_ptr_indir12; break;
1103     case 16: func = TOK___bound_ptr_indir16; break;
1104     default:
1105         tcc_error("unhandled size when dereferencing bounded pointer");
1106         func = 0;
1107         break;
1108     }
1109
1110     /* patch relocation */
1111     /* XXX: find a better solution ? */
1112     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1113     sym = external_global_sym(func, &func_old_type, 0);
1114     if (!sym->c)
1115         put_extern_sym(sym, NULL, 0, 0);
1116     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1117 }
1118 #endif
1119
1120 /* end of X86 code generator */
1121 /*************************************************************/
1122 #endif
1123 /*************************************************************/