i386-gen.c

   1 /*
   2  *  X86 code generator for TCC
   3  *
   4  *  Copyright (c) 2001-2004 Fabrice Bellard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, write to the Free Software
  18  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
  19  */
  20
  21 #ifdef TARGET_DEFS_ONLY
  22
  23 /* number of available registers */
  24 #define NB_REGS         4
  25 #define NB_ASM_REGS     8
  26
  27 /* a register can belong to several classes. The classes must be
  28    sorted from more general to more precise (see gv2() code which does
  29    assumptions on it). */
  30 #define RC_INT     0x0001 /* generic integer register */
  31 #define RC_FLOAT   0x0002 /* generic float register */
  32 #define RC_EAX     0x0004
  33 #define RC_ST0     0x0008
  34 #define RC_ECX     0x0010
  35 #define RC_EDX     0x0020
  36 #define RC_IRET    RC_EAX /* function return: integer register */
  37 #define RC_LRET    RC_EDX /* function return: second integer register */
  38 #define RC_FRET    RC_ST0 /* function return: float register */
  39
  40 /* pretty names for the registers */
  41 enum {
  42     TREG_EAX = 0,
  43     TREG_ECX,
  44     TREG_EDX,
  45     TREG_ST0,
  46     TREG_ESP = 4
  47 };
  48
  49 /* return registers for function */
  50 #define REG_IRET TREG_EAX /* single word int return register */
  51 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  52 #define REG_FRET TREG_ST0 /* float return register */
  53
  54 /* defined if function parameters must be evaluated in reverse order */
  55 #define INVERT_FUNC_PARAMS
  56
  57 /* defined if structures are passed as pointers. Otherwise structures
  58    are directly pushed on stack. */
  59 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  60
  61 /* pointer size, in bytes */
  62 #define PTR_SIZE 4
  63
  64 /* long double size and alignment, in bytes */
  65 #define LDOUBLE_SIZE  12
  66 #define LDOUBLE_ALIGN 4
  67 /* maximum alignment (for aligned attribute support) */
  68 #define MAX_ALIGN     8
  69
  70
  71 #define psym oad
  72
  73 /******************************************************/
  74 /* ELF defines */
  75
  76 #define EM_TCC_TARGET EM_386
  77
  78 /* relocation type for 32 bit data relocation */
  79 #define R_DATA_32   R_386_32
  80 #define R_DATA_PTR  R_386_32
  81 #define R_JMP_SLOT  R_386_JMP_SLOT
  82 #define R_COPY      R_386_COPY
  83
  84 #define ELF_START_ADDR 0x08048000
  85 #define ELF_PAGE_SIZE  0x1000
  86
  87 /******************************************************/
  88 #else /* ! TARGET_DEFS_ONLY */
  89 /******************************************************/
  90 #include "tcc.h"
  91
  92 ST_DATA const int reg_classes[NB_REGS] = {
  93     /* eax */ RC_INT | RC_EAX,
  94     /* ecx */ RC_INT | RC_ECX,
  95     /* edx */ RC_INT | RC_EDX,
  96     /* st0 */ RC_FLOAT | RC_ST0,
  97 };
  98
  99 static unsigned long func_sub_sp_offset;
 100 static int func_ret_sub;
 101 #ifdef CONFIG_TCC_BCHECK
 102 static addr_t func_bound_offset;
 103 #endif
 104
 105 /* XXX: make it faster ? */
 106 ST_FUNC void g(int c)
 107 {
 108     int ind1;
 109     ind1 = ind + 1;
 110     if (ind1 > cur_text_section->data_allocated)
 111         section_realloc(cur_text_section, ind1);
 112     cur_text_section->data[ind] = c;
 113     ind = ind1;
 114 }
 115
 116 ST_FUNC void o(unsigned int c)
 117 {
 118     while (c) {
 119         g(c);
 120         c = c >> 8;
 121     }
 122 }
 123
 124 ST_FUNC void gen_le16(int v)
 125 {
 126     g(v);
 127     g(v >> 8);
 128 }
 129
 130 ST_FUNC void gen_le32(int c)
 131 {
 132     g(c);
 133     g(c >> 8);
 134     g(c >> 16);
 135     g(c >> 24);
 136 }
 137
 138 /* output a symbol and patch all calls to it */
 139 ST_FUNC void gsym_addr(int t, int a)
 140 {
 141     while (t) {
 142         unsigned char *ptr = cur_text_section->data + t;
 143         uint32_t n = read32le(ptr); /* next value */
 144         write32le(ptr, a - t - 4);
 145         t = n;
 146     }
 147 }
 148
 149 ST_FUNC void gsym(int t)
 150 {
 151     gsym_addr(t, ind);
 152 }
 153
 154 /* psym is used to put an instruction with a data field which is a
 155    reference to a symbol. It is in fact the same as oad ! */
 156 #define psym oad
 157
 158 /* instruction + 4 bytes data. Return the address of the data */
 159 ST_FUNC int oad(int c, int s)
 160 {
 161     int ind1;
 162
 163     o(c);
 164     ind1 = ind + 4;
 165     if (ind1 > cur_text_section->data_allocated)
 166         section_realloc(cur_text_section, ind1);
 167     write32le(cur_text_section->data + ind, s);
 168     s = ind;
 169     ind = ind1;
 170     return s;
 171 }
 172
 173 /* output constant with relocation if 'r & VT_SYM' is true */
 174 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 175 {
 176     if (r & VT_SYM)
 177         greloc(cur_text_section, sym, ind, R_386_32);
 178     gen_le32(c);
 179 }
 180
 181 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 182 {
 183     if (r & VT_SYM)
 184         greloc(cur_text_section, sym, ind, R_386_PC32);
 185     gen_le32(c - 4);
 186 }
 187
 188 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 189    opcode bits */
 190 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 191 {
 192     op_reg = op_reg << 3;
 193     if ((r & VT_VALMASK) == VT_CONST) {
 194         /* constant memory reference */
 195         o(0x05 | op_reg);
 196         gen_addr32(r, sym, c);
 197     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 198         /* currently, we use only ebp as base */
 199         if (c == (char)c) {
 200             /* short reference */
 201             o(0x45 | op_reg);
 202             g(c);
 203         } else {
 204             oad(0x85 | op_reg, c);
 205         }
 206     } else {
 207         g(0x00 | op_reg | (r & VT_VALMASK));
 208     }
 209 }
 210
 211 /* load 'r' from value 'sv' */
 212 ST_FUNC void load(int r, SValue *sv)
 213 {
 214     int v, t, ft, fc, fr;
 215     SValue v1;
 216
 217 #ifdef TCC_TARGET_PE
 218     SValue v2;
 219     sv = pe_getimport(sv, &v2);
 220 #endif
 221
 222     fr = sv->r;
 223     ft = sv->type.t;
 224     fc = sv->c.i;
 225
 226     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.i = 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 || (ft & VT_TYPE) == VT_BOOL) {
 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.i;
 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 static void gen_static_call(int v)
 341 {
 342     Sym *sym;
 343
 344     sym = external_global_sym(v, &func_old_type, 0);
 345     oad(0xe8, -4);
 346     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 347 }
 348
 349 /* 'is_jmp' is '1' if it is a jump */
 350 static void gcall_or_jmp(int is_jmp)
 351 {
 352     int r;
 353     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 354         /* constant case */
 355         if (vtop->r & VT_SYM) {
 356             /* relocation case */
 357             greloc(cur_text_section, vtop->sym,
 358                    ind + 1, R_386_PC32);
 359         } else {
 360             /* put an empty PC32 relocation */
 361             put_elf_reloc(symtab_section, cur_text_section,
 362                           ind + 1, R_386_PC32, 0);
 363         }
 364         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 365     } else {
 366         /* otherwise, indirect call */
 367         r = gv(RC_INT);
 368         o(0xff); /* call/jmp *r */
 369         o(0xd0 + r + (is_jmp << 4));
 370     }
 371 }
 372
 373 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 374 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 375
 376 /* Return the number of registers needed to return the struct, or 0 if
 377    returning via struct pointer. */
 378 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 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     *regsize = 4;
 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     addr_t 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         addr_t saved_ind;
 595         addr_t *bounds_ptr;
 596         Sym *sym_data;
 597         /* add end of table info */
 598         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 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 = vtop->r & VT_VALMASK;
 674     if (v == VT_CMP) {
 675         /* fast case : can jump directly since flags are set */
 676         g(0x0f);
 677         t = psym((vtop->c.i - 16) ^ inv, t);
 678     } else if (v == VT_JMP || v == VT_JMPI) {
 679         /* && or || optimization */
 680         if ((v & 1) == inv) {
 681             uint32_t n1, n = vtop->c.i;
 682             /* insert vtop->c jump list in t */
 683             while ((n1 = read32le(cur_text_section->data + n)))
 684                 n = n1;
 685             write32le(cur_text_section->data + n, t);
 686             t = vtop->c.i;
 687         } else {
 688             t = gjmp(t);
 689             gsym(vtop->c.i);
 690         }
 691     }
 692     vtop--;
 693     return t;
 694 }
 695
 696 /* generate an integer binary operation */
 697 ST_FUNC void gen_opi(int op)
 698 {
 699     int r, fr, opc, c;
 700
 701     switch(op) {
 702     case '+':
 703     case TOK_ADDC1: /* add with carry generation */
 704         opc = 0;
 705     gen_op8:
 706         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 707             /* constant case */
 708             vswap();
 709             r = gv(RC_INT);
 710             vswap();
 711             c = vtop->c.i;
 712             if (c == (char)c) {
 713                 /* generate inc and dec for smaller code */
 714                 if (c==1 && opc==0) {
 715                     o (0x40 | r); // inc
 716                 } else if (c==1 && opc==5) {
 717                     o (0x48 | r); // dec
 718                 } else {
 719                     o(0x83);
 720                     o(0xc0 | (opc << 3) | r);
 721                     g(c);
 722                 }
 723             } else {
 724                 o(0x81);
 725                 oad(0xc0 | (opc << 3) | r, c);
 726             }
 727         } else {
 728             gv2(RC_INT, RC_INT);
 729             r = vtop[-1].r;
 730             fr = vtop[0].r;
 731             o((opc << 3) | 0x01);
 732             o(0xc0 + r + fr * 8);
 733         }
 734         vtop--;
 735         if (op >= TOK_ULT && op <= TOK_GT) {
 736             vtop->r = VT_CMP;
 737             vtop->c.i = op;
 738         }
 739         break;
 740     case '-':
 741     case TOK_SUBC1: /* sub with carry generation */
 742         opc = 5;
 743         goto gen_op8;
 744     case TOK_ADDC2: /* add with carry use */
 745         opc = 2;
 746         goto gen_op8;
 747     case TOK_SUBC2: /* sub with carry use */
 748         opc = 3;
 749         goto gen_op8;
 750     case '&':
 751         opc = 4;
 752         goto gen_op8;
 753     case '^':
 754         opc = 6;
 755         goto gen_op8;
 756     case '|':
 757         opc = 1;
 758         goto gen_op8;
 759     case '*':
 760         gv2(RC_INT, RC_INT);
 761         r = vtop[-1].r;
 762         fr = vtop[0].r;
 763         vtop--;
 764         o(0xaf0f); /* imul fr, r */
 765         o(0xc0 + fr + r * 8);
 766         break;
 767     case TOK_SHL:
 768         opc = 4;
 769         goto gen_shift;
 770     case TOK_SHR:
 771         opc = 5;
 772         goto gen_shift;
 773     case TOK_SAR:
 774         opc = 7;
 775     gen_shift:
 776         opc = 0xc0 | (opc << 3);
 777         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 778             /* constant case */
 779             vswap();
 780             r = gv(RC_INT);
 781             vswap();
 782             c = vtop->c.i & 0x1f;
 783             o(0xc1); /* shl/shr/sar $xxx, r */
 784             o(opc | r);
 785             g(c);
 786         } else {
 787             /* we generate the shift in ecx */
 788             gv2(RC_INT, RC_ECX);
 789             r = vtop[-1].r;
 790             o(0xd3); /* shl/shr/sar %cl, r */
 791             o(opc | r);
 792         }
 793         vtop--;
 794         break;
 795     case '/':
 796     case TOK_UDIV:
 797     case TOK_PDIV:
 798     case '%':
 799     case TOK_UMOD:
 800     case TOK_UMULL:
 801         /* first operand must be in eax */
 802         /* XXX: need better constraint for second operand */
 803         gv2(RC_EAX, RC_ECX);
 804         r = vtop[-1].r;
 805         fr = vtop[0].r;
 806         vtop--;
 807         save_reg(TREG_EDX);
 808         if (op == TOK_UMULL) {
 809             o(0xf7); /* mul fr */
 810             o(0xe0 + fr);
 811             vtop->r2 = TREG_EDX;
 812             r = TREG_EAX;
 813         } else {
 814             if (op == TOK_UDIV || op == TOK_UMOD) {
 815                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 816                 o(0xf0 + fr);
 817             } else {
 818                 o(0xf799); /* cltd, idiv fr, %eax */
 819                 o(0xf8 + fr);
 820             }
 821             if (op == '%' || op == TOK_UMOD)
 822                 r = TREG_EDX;
 823             else
 824                 r = TREG_EAX;
 825         }
 826         vtop->r = r;
 827         break;
 828     default:
 829         opc = 7;
 830         goto gen_op8;
 831     }
 832 }
 833
 834 /* generate a floating point operation 'v = t1 op t2' instruction. The
 835    two operands are guaranted to have the same floating point type */
 836 /* XXX: need to use ST1 too */
 837 ST_FUNC void gen_opf(int op)
 838 {
 839     int a, ft, fc, swapped, r;
 840
 841     /* convert constants to memory references */
 842     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 843         vswap();
 844         gv(RC_FLOAT);
 845         vswap();
 846     }
 847     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 848         gv(RC_FLOAT);
 849
 850     /* must put at least one value in the floating point register */
 851     if ((vtop[-1].r & VT_LVAL) &&
 852         (vtop[0].r & VT_LVAL)) {
 853         vswap();
 854         gv(RC_FLOAT);
 855         vswap();
 856     }
 857     swapped = 0;
 858     /* swap the stack if needed so that t1 is the register and t2 is
 859        the memory reference */
 860     if (vtop[-1].r & VT_LVAL) {
 861         vswap();
 862         swapped = 1;
 863     }
 864     if (op >= TOK_ULT && op <= TOK_GT) {
 865         /* load on stack second operand */
 866         load(TREG_ST0, vtop);
 867         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 868         if (op == TOK_GE || op == TOK_GT)
 869             swapped = !swapped;
 870         else if (op == TOK_EQ || op == TOK_NE)
 871             swapped = 0;
 872         if (swapped)
 873             o(0xc9d9); /* fxch %st(1) */
 874         if (op == TOK_EQ || op == TOK_NE)
 875             o(0xe9da); /* fucompp */
 876         else
 877             o(0xd9de); /* fcompp */
 878         o(0xe0df); /* fnstsw %ax */
 879         if (op == TOK_EQ) {
 880             o(0x45e480); /* and $0x45, %ah */
 881             o(0x40fC80); /* cmp $0x40, %ah */
 882         } else if (op == TOK_NE) {
 883             o(0x45e480); /* and $0x45, %ah */
 884             o(0x40f480); /* xor $0x40, %ah */
 885             op = TOK_NE;
 886         } else if (op == TOK_GE || op == TOK_LE) {
 887             o(0x05c4f6); /* test $0x05, %ah */
 888             op = TOK_EQ;
 889         } else {
 890             o(0x45c4f6); /* test $0x45, %ah */
 891             op = TOK_EQ;
 892         }
 893         vtop--;
 894         vtop->r = VT_CMP;
 895         vtop->c.i = op;
 896     } else {
 897         /* no memory reference possible for long double operations */
 898         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 899             load(TREG_ST0, vtop);
 900             swapped = !swapped;
 901         }
 902
 903         switch(op) {
 904         default:
 905         case '+':
 906             a = 0;
 907             break;
 908         case '-':
 909             a = 4;
 910             if (swapped)
 911                 a++;
 912             break;
 913         case '*':
 914             a = 1;
 915             break;
 916         case '/':
 917             a = 6;
 918             if (swapped)
 919                 a++;
 920             break;
 921         }
 922         ft = vtop->type.t;
 923         fc = vtop->c.i;
 924         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 925             o(0xde); /* fxxxp %st, %st(1) */
 926             o(0xc1 + (a << 3));
 927         } else {
 928             /* if saved lvalue, then we must reload it */
 929             r = vtop->r;
 930             if ((r & VT_VALMASK) == VT_LLOCAL) {
 931                 SValue v1;
 932                 r = get_reg(RC_INT);
 933                 v1.type.t = VT_INT;
 934                 v1.r = VT_LOCAL | VT_LVAL;
 935                 v1.c.i = fc;
 936                 load(r, &v1);
 937                 fc = 0;
 938             }
 939
 940             if ((ft & VT_BTYPE) == VT_DOUBLE)
 941                 o(0xdc);
 942             else
 943                 o(0xd8);
 944             gen_modrm(a, r, vtop->sym, fc);
 945         }
 946         vtop--;
 947     }
 948 }
 949
 950 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
 951    and 'long long' cases. */
 952 ST_FUNC void gen_cvt_itof(int t)
 953 {
 954     save_reg(TREG_ST0);
 955     gv(RC_INT);
 956     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 957         /* signed long long to float/double/long double (unsigned case
 958            is handled generically) */
 959         o(0x50 + vtop->r2); /* push r2 */
 960         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 961         o(0x242cdf); /* fildll (%esp) */
 962         o(0x08c483); /* add $8, %esp */
 963     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
 964                (VT_INT | VT_UNSIGNED)) {
 965         /* unsigned int to float/double/long double */
 966         o(0x6a); /* push $0 */
 967         g(0x00);
 968         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 969         o(0x242cdf); /* fildll (%esp) */
 970         o(0x08c483); /* add $8, %esp */
 971     } else {
 972         /* int to float/double/long double */
 973         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
 974         o(0x2404db); /* fildl (%esp) */
 975         o(0x04c483); /* add $4, %esp */
 976     }
 977     vtop->r = TREG_ST0;
 978 }
 979
 980 /* convert fp to int 't' type */
 981 ST_FUNC void gen_cvt_ftoi(int t)
 982 {
 983  #ifndef COMMIT_4ad186c5ef61_IS_FIXED
 984     /* a good version but it takes a more time to execute */
 985     gv(RC_FLOAT);
 986     save_reg(TREG_EAX);
 987     save_reg(TREG_EDX);
 988     gen_static_call(TOK___tcc_cvt_ftol);
 989     vtop->r = TREG_EAX; /* mark reg as used */
 990     if (t == VT_LLONG)
 991         vtop->r2 = TREG_EDX;
 992  #else
 993     /* a new version with a bug: t2a = 44100312 */
 994     /*
 995     #include<stdio.h>
 996     int main() {
 997         int t1 = 176401255;
 998         float f = 0.25f;
 999         int t2a = (int)(t1 * f); // must be 44100313
1000         int t2b = (int)(t1 * (float)0.25f);
1001         printf("t2a=%d t2b=%d \n",t2a,t2b);
1002         return 0;
1003     }
1004     */
1005     int bt = vtop->type.t & VT_BTYPE;
1006     if (bt == VT_FLOAT)
1007         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1008     else if (bt == VT_LDOUBLE)
1009         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1010     else
1011         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1012     vswap();
1013     gfunc_call(1);
1014     vpushi(0);
1015     vtop->r = REG_IRET;
1016     vtop->r2 = REG_LRET;
1017  #endif
1018 }
1019
1020 /* convert from one floating point type to another */
1021 ST_FUNC void gen_cvt_ftof(int t)
1022 {
1023     /* all we have to do on i386 is to put the float in a register */
1024     gv(RC_FLOAT);
1025 }
1026
1027 /* computed goto support */
1028 ST_FUNC void ggoto(void)
1029 {
1030     gcall_or_jmp(1);
1031     vtop--;
1032 }
1033
1034 /* bound check support functions */
1035 #ifdef CONFIG_TCC_BCHECK
1036
1037 /* generate a bounded pointer addition */
1038 ST_FUNC void gen_bounded_ptr_add(void)
1039 {
1040     /* prepare fast i386 function call (args in eax and edx) */
1041     gv2(RC_EAX, RC_EDX);
1042     /* save all temporary registers */
1043     vtop -= 2;
1044     save_regs(0);
1045     /* do a fast function call */
1046     gen_static_call(TOK___bound_ptr_add);
1047     /* returned pointer is in eax */
1048     vtop++;
1049     vtop->r = TREG_EAX | VT_BOUNDED;
1050     /* address of bounding function call point */
1051     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1052 }
1053
1054 /* patch pointer addition in vtop so that pointer dereferencing is
1055    also tested */
1056 ST_FUNC void gen_bounded_ptr_deref(void)
1057 {
1058     addr_t func;
1059     addr_t size, align;
1060     Elf32_Rel *rel;
1061     Sym *sym;
1062
1063     size = 0;
1064     /* XXX: put that code in generic part of tcc */
1065     if (!is_float(vtop->type.t)) {
1066         if (vtop->r & VT_LVAL_BYTE)
1067             size = 1;
1068         else if (vtop->r & VT_LVAL_SHORT)
1069             size = 2;
1070     }
1071     if (!size)
1072         size = type_size(&vtop->type, &align);
1073     switch(size) {
1074     case  1: func = TOK___bound_ptr_indir1; break;
1075     case  2: func = TOK___bound_ptr_indir2; break;
1076     case  4: func = TOK___bound_ptr_indir4; break;
1077     case  8: func = TOK___bound_ptr_indir8; break;
1078     case 12: func = TOK___bound_ptr_indir12; break;
1079     case 16: func = TOK___bound_ptr_indir16; break;
1080     default:
1081         tcc_error("unhandled size when dereferencing bounded pointer");
1082         func = 0;
1083         break;
1084     }
1085
1086     /* patch relocation */
1087     /* XXX: find a better solution ? */
1088     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1089     sym = external_global_sym(func, &func_old_type, 0);
1090     if (!sym->c)
1091         put_extern_sym(sym, NULL, 0, 0);
1092     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1093 }
1094 #endif
1095
1096 /* Save the stack pointer onto the stack */
1097 ST_FUNC void gen_vla_sp_save(int addr) {
1098     /* mov %esp,addr(%ebp)*/
1099     o(0x89);
1100     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1101 }
1102
1103 /* Restore the SP from a location on the stack */
1104 ST_FUNC void gen_vla_sp_restore(int addr) {
1105     o(0x8b);
1106     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1107 }
1108
1109 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1110 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1111 #ifdef TCC_TARGET_PE
1112     /* alloca does more than just adjust %rsp on Windows */
1113     vpush_global_sym(&func_old_type, TOK_alloca);
1114     vswap(); /* Move alloca ref past allocation size */
1115     gfunc_call(1);
1116 #else
1117     int r;
1118     r = gv(RC_INT); /* allocation size */
1119     /* sub r,%rsp */
1120     o(0x2b);
1121     o(0xe0 | r);
1122     /* We align to 16 bytes rather than align */
1123     /* and ~15, %esp */
1124     o(0xf0e483);
1125     vpop();
1126 #endif
1127 }
1128
1129 /* end of X86 code generator */
1130 /*************************************************************/
1131 #endif
1132 /*************************************************************/