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         8
  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_EBX     0x0040
  37 #define RC_ESI     0x0080
  38 #define RC_EDI     0x0100
  39 #define RC_INT2     0x0200
  40 #define RC_IRET    RC_EAX /* function return: integer register */
  41 #define RC_LRET    RC_EDX /* function return: second integer register */
  42 #define RC_FRET    RC_ST0 /* function return: float register */
  43 #define RC_MASK    (RC_INT|RC_INT2|RC_FLOAT)
  44 /* pretty names for the registers */
  45 enum {
  46     TREG_EAX = 0,
  47     TREG_ECX,
  48     TREG_EDX,
  49     TREG_EBX,
  50     TREG_ESP,
  51     TREG_ST0,
  52     TREG_ESI,
  53     TREG_EDI,
  54 };
  55
  56 /* return registers for function */
  57 #define REG_IRET TREG_EAX /* single word int return register */
  58 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  59 #define REG_FRET TREG_ST0 /* float return register */
  60
  61 /* defined if function parameters must be evaluated in reverse order */
  62 #define INVERT_FUNC_PARAMS
  63
  64 /* defined if structures are passed as pointers. Otherwise structures
  65    are directly pushed on stack. */
  66 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  67
  68 /* pointer size, in bytes */
  69 #define PTR_SIZE 4
  70
  71 /* long double size and alignment, in bytes */
  72 #define LDOUBLE_SIZE  12
  73 #define LDOUBLE_ALIGN 4
  74 /* maximum alignment (for aligned attribute support) */
  75 #define MAX_ALIGN     8
  76
  77
  78 #define psym oad
  79
  80 /******************************************************/
  81 /* ELF defines */
  82
  83 #define EM_TCC_TARGET EM_386
  84
  85 /* relocation type for 32 bit data relocation */
  86 #define R_DATA_32   R_386_32
  87 #define R_DATA_PTR  R_386_32
  88 #define R_JMP_SLOT  R_386_JMP_SLOT
  89 #define R_COPY      R_386_COPY
  90
  91 #define ELF_START_ADDR 0x08048000
  92 #define ELF_PAGE_SIZE  0x1000
  93
  94 /******************************************************/
  95 #else /* ! TARGET_DEFS_ONLY */
  96 /******************************************************/
  97 #include "tcc.h"
  98
  99 ST_DATA const int reg_classes[NB_REGS] = {
 100     /* eax */ RC_INT | RC_EAX | RC_INT2,
 101     /* ecx */ RC_INT | RC_ECX | RC_INT2,
 102     /* edx */ RC_INT | RC_EDX,
 103     RC_INT|RC_INT2|RC_EBX,
 104     0,
 105     /* st0 */ RC_FLOAT | RC_ST0,
 106     RC_ESI|RC_INT2,
 107     RC_EDI|RC_INT2,
 108 };
 109
 110 static unsigned long func_sub_sp_offset;
 111 static int func_ret_sub;
 112 #ifdef CONFIG_TCC_BCHECK
 113 static unsigned long func_bound_offset;
 114 #endif
 115
 116 /* XXX: make it faster ? */
 117 ST_FUNC void g(int c)
 118 {
 119     int ind1;
 120     ind1 = ind + 1;
 121     if (ind1 > cur_text_section->data_allocated)
 122         section_realloc(cur_text_section, ind1);
 123     cur_text_section->data[ind] = c;
 124     ind = ind1;
 125 }
 126
 127 ST_FUNC void o(unsigned int c)
 128 {
 129     while (c) {
 130         g(c);
 131         c = c >> 8;
 132     }
 133 }
 134
 135 ST_FUNC void gen_le16(int v)
 136 {
 137     g(v);
 138     g(v >> 8);
 139 }
 140
 141 ST_FUNC void gen_le32(int c)
 142 {
 143     g(c);
 144     g(c >> 8);
 145     g(c >> 16);
 146     g(c >> 24);
 147 }
 148
 149 /* output a symbol and patch all calls to it */
 150 ST_FUNC void gsym_addr(int t, int a)
 151 {
 152     int n, *ptr;
 153     while (t) {
 154         ptr = (int *)(cur_text_section->data + t);
 155         n = *ptr; /* next value */
 156         *ptr = a - t - 4;
 157         t = n;
 158     }
 159 }
 160
 161 ST_FUNC void gsym(int t)
 162 {
 163     gsym_addr(t, ind);
 164 }
 165
 166 /* psym is used to put an instruction with a data field which is a
 167    reference to a symbol. It is in fact the same as oad ! */
 168 #define psym oad
 169
 170 /* instruction + 4 bytes data. Return the address of the data */
 171 ST_FUNC int oad(int c, int s)
 172 {
 173     int ind1;
 174
 175     o(c);
 176     ind1 = ind + 4;
 177     if (ind1 > cur_text_section->data_allocated)
 178         section_realloc(cur_text_section, ind1);
 179     *(int *)(cur_text_section->data + ind) = s;
 180     s = ind;
 181     ind = ind1;
 182     return s;
 183 }
 184
 185 /* output constant with relocation if 'r & VT_SYM' is true */
 186 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 187 {
 188     if (r & VT_SYM)
 189         greloc(cur_text_section, sym, ind, R_386_32);
 190     gen_le32(c);
 191 }
 192
 193 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 194 {
 195     if (r & VT_SYM)
 196         greloc(cur_text_section, sym, ind, R_386_PC32);
 197     gen_le32(c - 4);
 198 }
 199
 200 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 201    opcode bits */
 202 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 203 {
 204     op_reg = op_reg << 3;
 205     if ((r & VT_VALMASK) == VT_CONST) {
 206         /* constant memory reference */
 207         o(0x05 | op_reg);
 208         gen_addr32(r, sym, c);
 209     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 210         /* currently, we use only ebp as base */
 211         if (c == (char)c) {
 212             /* short reference */
 213             o(0x45 | op_reg);
 214             g(c);
 215         } else {
 216             oad(0x85 | op_reg, c);
 217         }
 218     } else {
 219         g(0x00 | op_reg | (r & VT_VALMASK));
 220     }
 221 }
 222
 223 /* load 'r' from value 'sv' */
 224 ST_FUNC void load(int r, SValue *sv)
 225 {
 226     int v, t, ft, fc, fr;
 227     SValue v1;
 228
 229 #ifdef TCC_TARGET_PE
 230     SValue v2;
 231     sv = pe_getimport(sv, &v2);
 232 #endif
 233
 234     fr = sv->r;
 235     ft = sv->type.t;
 236     fc = sv->c.ul;
 237
 238     v = fr & VT_VALMASK;
 239     if (fr & VT_LVAL) {
 240         if(fr & VT_TMP){
 241                         int size, align;
 242                         if((ft & VT_BTYPE) == VT_FUNC)
 243                                 size = PTR_SIZE;
 244                         else
 245                                 size = type_size(&sv->type, &align);
 246                         loc_stack(size, 0);
 247                 }
 248         if (v == VT_LLOCAL) {
 249             v1.type.t = VT_INT;
 250             v1.r = VT_LOCAL | VT_LVAL;
 251             v1.c.ul = fc;
 252             fr = r;
 253             if (!(reg_classes[fr] & RC_INT))
 254                 fr = get_reg(RC_INT);
 255             load(fr, &v1);
 256             fc = 0;
 257         }
 258         if ((ft & VT_BTYPE) == VT_FLOAT) {
 259             o(0xd9); /* flds */
 260             r = 0;
 261         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 262             o(0xdd); /* fldl */
 263             r = 0;
 264         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 265             o(0xdb); /* fldt */
 266             r = 5;
 267         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 268             o(0xbe0f);   /* movsbl */
 269         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 270             o(0xb60f);   /* movzbl */
 271         } else if ((ft & VT_TYPE) == VT_SHORT) {
 272             o(0xbf0f);   /* movswl */
 273         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 274             o(0xb70f);   /* movzwl */
 275         } else {
 276             o(0x8b);     /* movl */
 277         }
 278         gen_modrm(r, fr, sv->sym, fc);
 279     } else {
 280         if (v == VT_CONST) {
 281             o(0xb8 + r); /* mov $xx, r */
 282             gen_addr32(fr, sv->sym, fc);
 283         } else if (v == VT_LOCAL) {
 284             if (fc) {
 285                 o(0x8d); /* lea xxx(%ebp), r */
 286                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 287             } else {
 288                 o(0x89);
 289                 o(0xe8 + r); /* mov %ebp, r */
 290             }
 291         } else if (v == VT_CMP) {
 292             oad(0xb8 + r, 0); /* mov $0, r */
 293             o(0x0f); /* setxx %br */
 294             o(fc);
 295             o(0xc0 + r);
 296         } else if (v == VT_JMP || v == VT_JMPI) {
 297             t = v & 1;
 298             oad(0xb8 + r, t); /* mov $1, r */
 299             o(0x05eb); /* jmp after */
 300             gsym(fc);
 301             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 302         } else if (v != r) {
 303             o(0x89);
 304             o(0xc0 + r + v * 8); /* mov v, r */
 305         }
 306     }
 307 }
 308
 309 /* store register 'r' in lvalue 'v' */
 310 ST_FUNC void store(int r, SValue *v)
 311 {
 312     int fr, bt, ft, fc;
 313
 314 #ifdef TCC_TARGET_PE
 315     SValue v2;
 316     v = pe_getimport(v, &v2);
 317 #endif
 318
 319     ft = v->type.t;
 320     fc = v->c.ul;
 321     fr = v->r & VT_VALMASK;
 322     bt = ft & VT_BTYPE;
 323     /* XXX: incorrect if float reg to reg */
 324     if (bt == VT_FLOAT) {
 325         o(0xd9); /* fsts */
 326         r = 2;
 327     } else if (bt == VT_DOUBLE) {
 328         o(0xdd); /* fstpl */
 329         r = 2;
 330     } else if (bt == VT_LDOUBLE) {
 331         o(0xc0d9); /* fld %st(0) */
 332         o(0xdb); /* fstpt */
 333         r = 7;
 334     } else {
 335         if (bt == VT_SHORT)
 336             o(0x66);
 337         if (bt == VT_BYTE || bt == VT_BOOL)
 338             o(0x88);
 339         else
 340             o(0x89);
 341     }
 342     if (fr == VT_CONST ||
 343         fr == VT_LOCAL ||
 344         (v->r & VT_LVAL)) {
 345         gen_modrm(r, v->r, v->sym, fc);
 346     } else if (fr != r) {
 347         o(0xc0 + fr + r * 8); /* mov r, fr */
 348     }
 349 }
 350
 351 static void gadd_sp(int val)
 352 {
 353     if (val == (char)val) {
 354         o(0xc483);
 355         g(val);
 356     } else {
 357         oad(0xc481, val); /* add $xxx, %esp */
 358     }
 359 }
 360
 361 static void gen_static_call(int v)
 362 {
 363     Sym *sym;
 364
 365     sym = external_global_sym(v, &func_old_type, 0);
 366     oad(0xe8, -4);
 367     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 368 }
 369
 370 /* 'is_jmp' is '1' if it is a jump */
 371 static void gcall_or_jmp(int is_jmp)
 372 {
 373     int r;
 374     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 375         /* constant case */
 376         if (vtop->r & VT_SYM) {
 377             /* relocation case */
 378             greloc(cur_text_section, vtop->sym,
 379                    ind + 1, R_386_PC32);
 380         } else {
 381             /* put an empty PC32 relocation */
 382             put_elf_reloc(symtab_section, cur_text_section,
 383                           ind + 1, R_386_PC32, 0);
 384         }
 385         oad(0xe8 + is_jmp, vtop->c.ul - 4); /* call/jmp im */
 386     } else {
 387         /* otherwise, indirect call */
 388         r = gv(RC_INT);
 389         o(0xff); /* call/jmp *r */
 390         o(0xd0 + r + (is_jmp << 4));
 391     }
 392 }
 393
 394 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 395 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 396
 397 /* Return the number of registers needed to return the struct, or 0 if
 398    returning via struct pointer. */
 399 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align)
 400 {
 401 #ifdef TCC_TARGET_PE
 402     int size, align;
 403
 404     *ret_align = 1; // Never have to re-align return values for x86
 405     size = type_size(vt, &align);
 406     if (size > 8) {
 407         return 0;
 408     } else if (size > 4) {
 409         ret->ref = NULL;
 410         ret->t = VT_LLONG;
 411         return 1;
 412     } else {
 413         ret->ref = NULL;
 414         ret->t = VT_INT;
 415         return 1;
 416     }
 417 #else
 418     *ret_align = 1; // Never have to re-align return values for x86
 419     return 0;
 420 #endif
 421 }
 422
 423 /* Generate function call. The function address is pushed first, then
 424    all the parameters in call order. This functions pops all the
 425    parameters and the function address. */
 426 ST_FUNC void gfunc_call(int nb_args)
 427 {
 428     int size, align, r, args_size, i, func_call;
 429     Sym *func_sym;
 430
 431     args_size = 0;
 432     for(i = 0;i < nb_args; i++) {
 433         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 434             size = type_size(&vtop->type, &align);
 435             /* align to stack align size */
 436             size = (size + 3) & ~3;
 437             /* allocate the necessary size on stack */
 438             oad(0xec81, size); /* sub $xxx, %esp */
 439             /* generate structure store */
 440             r = get_reg(RC_INT);
 441             o(0x89); /* mov %esp, r */
 442             o(0xe0 + r);
 443             vset(&vtop->type, r | VT_LVAL, 0);
 444             vswap();
 445             vstore();
 446             args_size += size;
 447         } else if (is_float(vtop->type.t)) {
 448             gv(RC_FLOAT); /* only one float register */
 449             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 450                 size = 4;
 451             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 452                 size = 8;
 453             else
 454                 size = 12;
 455             oad(0xec81, size); /* sub $xxx, %esp */
 456             if (size == 12)
 457                 o(0x7cdb);
 458             else
 459                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 460             g(0x24);
 461             g(0x00);
 462             args_size += size;
 463         } else {
 464             /* simple type (currently always same size) */
 465             /* XXX: implicit cast ? */
 466             r = gv(RC_INT);
 467             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 468                 size = 8;
 469                 o(0x50 + vtop->r2); /* push r */
 470             } else {
 471                 size = 4;
 472             }
 473             o(0x50 + r); /* push r */
 474             args_size += size;
 475         }
 476         vtop--;
 477     }
 478     save_regs(0); /* save used temporary registers */
 479     func_sym = vtop->type.ref;
 480     func_call = func_sym->a.func_call;
 481     /* fast call case */
 482     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 483         func_call == FUNC_FASTCALLW) {
 484         int fastcall_nb_regs;
 485         uint8_t *fastcall_regs_ptr;
 486         if (func_call == FUNC_FASTCALLW) {
 487             fastcall_regs_ptr = fastcallw_regs;
 488             fastcall_nb_regs = 2;
 489         } else {
 490             fastcall_regs_ptr = fastcall_regs;
 491             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 492         }
 493         for(i = 0;i < fastcall_nb_regs; i++) {
 494             if (args_size <= 0)
 495                 break;
 496             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 497             /* XXX: incorrect for struct/floats */
 498             args_size -= 4;
 499         }
 500     }
 501 #ifndef TCC_TARGET_PE
 502     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 503         args_size -= 4;
 504 #endif
 505     gcall_or_jmp(0);
 506
 507     if (args_size && func_call != FUNC_STDCALL)
 508         gadd_sp(args_size);
 509     vtop--;
 510 }
 511
 512 #ifdef TCC_TARGET_PE
 513 #define FUNC_PROLOG_SIZE 10
 514 #else
 515 #define FUNC_PROLOG_SIZE 9
 516 #endif
 517
 518 /* generate function prolog of type 't' */
 519 ST_FUNC void gfunc_prolog(CType *func_type)
 520 {
 521     int addr, align, size, func_call, fastcall_nb_regs;
 522     int param_index, param_addr;
 523     uint8_t *fastcall_regs_ptr;
 524     Sym *sym;
 525     CType *type;
 526
 527     sym = func_type->ref;
 528     func_call = sym->a.func_call;
 529     addr = 8;
 530     loc = 0;
 531     func_vc = 0;
 532
 533     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 534         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 535         fastcall_regs_ptr = fastcall_regs;
 536     } else if (func_call == FUNC_FASTCALLW) {
 537         fastcall_nb_regs = 2;
 538         fastcall_regs_ptr = fastcallw_regs;
 539     } else {
 540         fastcall_nb_regs = 0;
 541         fastcall_regs_ptr = NULL;
 542     }
 543     param_index = 0;
 544
 545     ind += FUNC_PROLOG_SIZE;
 546     func_sub_sp_offset = ind;
 547     /* if the function returns a structure, then add an
 548        implicit pointer parameter */
 549     func_vt = sym->type;
 550     func_var = (sym->c == FUNC_ELLIPSIS);
 551 #ifdef TCC_TARGET_PE
 552     size = type_size(&func_vt,&align);
 553     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 554 #else
 555     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 556 #endif
 557         /* XXX: fastcall case ? */
 558         func_vc = addr;
 559         addr += 4;
 560         param_index++;
 561     }
 562     /* define parameters */
 563     while ((sym = sym->next) != NULL) {
 564         type = &sym->type;
 565         size = type_size(type, &align);
 566         size = (size + 3) & ~3;
 567 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 568         /* structs are passed as pointer */
 569         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 570             size = 4;
 571         }
 572 #endif
 573         if (param_index < fastcall_nb_regs) {
 574             /* save FASTCALL register */
 575             loc -= 4;
 576             o(0x89);     /* movl */
 577             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 578             param_addr = loc;
 579         } else {
 580             param_addr = addr;
 581             addr += size;
 582         }
 583         sym_push(sym->v & ~SYM_FIELD, type,
 584                  VT_LOCAL | lvalue_type(type->t), param_addr);
 585         param_index++;
 586     }
 587     func_ret_sub = 0;
 588     /* pascal type call ? */
 589     if (func_call == FUNC_STDCALL)
 590         func_ret_sub = addr - 8;
 591 #ifndef TCC_TARGET_PE
 592     else if (func_vc)
 593         func_ret_sub = 4;
 594 #endif
 595
 596 #ifdef CONFIG_TCC_BCHECK
 597     /* leave some room for bound checking code */
 598     if (tcc_state->do_bounds_check) {
 599         oad(0xb8, 0); /* lbound section pointer */
 600         oad(0xb8, 0); /* call to function */
 601         func_bound_offset = lbounds_section->data_offset;
 602     }
 603 #endif
 604 }
 605
 606 /* generate function epilog */
 607 ST_FUNC void gfunc_epilog(void)
 608 {
 609     int v, saved_ind;
 610
 611 #ifdef CONFIG_TCC_BCHECK
 612     if (tcc_state->do_bounds_check
 613      && func_bound_offset != lbounds_section->data_offset) {
 614         int saved_ind;
 615         int *bounds_ptr;
 616         Sym *sym_data;
 617         /* add end of table info */
 618         bounds_ptr = section_ptr_add(lbounds_section, sizeof(int));
 619         *bounds_ptr = 0;
 620         /* generate bound local allocation */
 621         saved_ind = ind;
 622         ind = func_sub_sp_offset;
 623         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 624                                func_bound_offset, lbounds_section->data_offset);
 625         greloc(cur_text_section, sym_data,
 626                ind + 1, R_386_32);
 627         oad(0xb8, 0); /* mov %eax, xxx */
 628         gen_static_call(TOK___bound_local_new);
 629
 630         ind = saved_ind;
 631         /* generate bound check local freeing */
 632         o(0x5250); /* save returned value, if any */
 633         greloc(cur_text_section, sym_data,
 634                ind + 1, R_386_32);
 635         oad(0xb8, 0); /* mov %eax, xxx */
 636         gen_static_call(TOK___bound_local_delete);
 637
 638         o(0x585a); /* restore returned value, if any */
 639     }
 640 #endif
 641     o(0xc9); /* leave */
 642     if (func_ret_sub == 0) {
 643         o(0xc3); /* ret */
 644     } else {
 645         o(0xc2); /* ret n */
 646         g(func_ret_sub);
 647         g(func_ret_sub >> 8);
 648     }
 649     /* align local size to word & save local variables */
 650
 651     v = (-loc + 3) & -4;
 652     saved_ind = ind;
 653     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 654 #ifdef TCC_TARGET_PE
 655     if (v >= 4096) {
 656         oad(0xb8, v); /* mov stacksize, %eax */
 657         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 658     } else
 659 #endif
 660     {
 661         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 662         o(0xec81);  /* sub esp, stacksize */
 663         gen_le32(v);
 664 #if FUNC_PROLOG_SIZE == 10
 665         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 666 #endif
 667     }
 668     ind = saved_ind;
 669 }
 670
 671 /* generate a jump to a label */
 672 ST_FUNC int gjmp(int t)
 673 {
 674     return psym(0xe9, t);
 675 }
 676
 677 /* generate a jump to a fixed address */
 678 ST_FUNC void gjmp_addr(int a)
 679 {
 680     int r;
 681     r = a - ind - 2;
 682     if (r == (char)r) {
 683         g(0xeb);
 684         g(r);
 685     } else {
 686         oad(0xe9, a - ind - 5);
 687     }
 688 }
 689
 690 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 691 ST_FUNC int gtst(int inv, int t)
 692 {
 693     int v, *p;
 694
 695     v = vtop->r & VT_VALMASK;
 696     if (v == VT_CMP) {
 697         /* fast case : can jump directly since flags are set */
 698         g(0x0f);
 699         t = psym((vtop->c.i - 16) ^ inv, t);
 700     } else if (v == VT_JMP || v == VT_JMPI) {
 701         /* && or || optimization */
 702         if ((v & 1) == inv) {
 703             /* insert vtop->c jump list in t */
 704             p = &vtop->c.i;
 705             while (*p != 0)
 706                 p = (int *)(cur_text_section->data + *p);
 707             *p = t;
 708             t = vtop->c.i;
 709         } else {
 710             t = gjmp(t);
 711             gsym(vtop->c.i);
 712         }
 713         } else {
 714         if (is_float(vtop->type.t) ||
 715             (vtop->type.t & VT_BTYPE) == VT_LLONG) {
 716             vpushi(0);
 717             gen_op(TOK_NE);
 718         }
 719         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 720             /* constant jmp optimization */
 721             if ((vtop->c.i != 0) != inv)
 722                 t = gjmp(t);
 723         } else {
 724             v = gv(RC_INT);
 725             o(0x85);
 726             o(0xc0 + v * 9);
 727             g(0x0f);
 728             t = psym(0x85 ^ inv, t);
 729         }
 730     }
 731     vtop--;
 732     return t;
 733 }
 734
 735 /* generate an integer binary operation */
 736 ST_FUNC void gen_opi(int op)
 737 {
 738     int r, fr, opc, fc, c;
 739         int cc, uu, tt2;
 740
 741         fr = vtop[0].r;
 742         fc = vtop->c.ul;
 743         cc = (fr & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST;
 744         tt2 = (fr & (VT_LVAL | VT_LVAL_TYPE)) == VT_LVAL;
 745
 746     switch(op) {
 747     case '+':
 748     case TOK_ADDC1: /* add with carry generation */
 749         opc = 0;
 750     gen_op8:
 751                 vswap();
 752                 r = gv(RC_INT);
 753                 vswap();
 754         if (cc) {
 755             /* constant case */
 756             c = vtop->c.i;
 757             if (c == (char)c) {
 758                 /* generate inc and dec for smaller code */
 759                 if (c == 1 && opc == 0) {
 760                     o (0x40 | r); // inc
 761                 } else if (c == 1 && opc == 5) {
 762                     o (0x48 | r); // dec
 763                 } else {
 764                     o(0x83);
 765                     o(0xc0 + r + opc*8);
 766                     g(c);
 767                 }
 768             } else {
 769                 o(0x81);
 770                 oad(0xc0 + r+ opc*8, c);
 771             }
 772         } else {
 773                         if(!tt2)
 774                                 fr = gv(RC_INT);
 775                         o(0x03 + opc*8);
 776                         if(fr >= VT_CONST)
 777                 gen_modrm(r, fr, vtop->sym, fc);
 778                         else
 779                                 o(0xc0 + fr + r*8);
 780         }
 781         vtop--;
 782         if (op >= TOK_ULT && op <= TOK_GT) {
 783             vtop->r = VT_CMP;
 784             vtop->c.i = op;
 785         }
 786         break;
 787     case '-':
 788     case TOK_SUBC1: /* sub with carry generation */
 789         opc = 5;
 790         goto gen_op8;
 791     case TOK_ADDC2: /* add with carry use */
 792         opc = 2;
 793         goto gen_op8;
 794     case TOK_SUBC2: /* sub with carry use */
 795         opc = 3;
 796         goto gen_op8;
 797     case '&':
 798         opc = 4;
 799         goto gen_op8;
 800     case '^':
 801         opc = 6;
 802         goto gen_op8;
 803     case '|':
 804         opc = 1;
 805         goto gen_op8;
 806     case '*':
 807                 opc = 5;
 808                 vswap();
 809                 r = gv(RC_INT);
 810                 vswap();
 811                 if(!tt2)
 812                         fr = gv(RC_INT);
 813                 if(r == TREG_EAX){
 814                         if(fr != TREG_EDX)
 815                                 save_reg(TREG_EDX);
 816                         o(0xf7);
 817                         if(fr >= VT_CONST)
 818                                 gen_modrm(opc, fr, vtop->sym, fc);
 819                         else
 820                                 o(0xc0 + fr + opc*8);
 821                 }else{
 822                         o(0xaf0f); /* imul fr, r */
 823                         if(fr >= VT_CONST)
 824                                 gen_modrm(r, fr, vtop->sym, fc);
 825                         else
 826                                 o(0xc0 + fr + r*8);
 827                 }
 828         vtop--;
 829         break;
 830     case TOK_SHL:
 831         opc = 4;
 832         goto gen_shift;
 833     case TOK_SHR:
 834         opc = 5;
 835         goto gen_shift;
 836     case TOK_SAR:
 837         opc = 7;
 838     gen_shift:
 839                 if (cc) {
 840             /* constant case */
 841             vswap();
 842             r = gv(RC_INT);
 843             vswap();
 844                         c = vtop->c.i;
 845                         if(c == 1){
 846                                 o(0xd1);
 847                                 o(0xc0 + r + opc*8);
 848                         }else{
 849                                 o(0xc1); /* shl/shr/sar $xxx, r */
 850                                 o(0xc0 + r + opc*8);
 851                                 g(c & 0x1f);
 852                         }
 853         } else {
 854             /* we generate the shift in ecx */
 855             gv2(RC_INT, RC_ECX);
 856             r = vtop[-1].r;
 857             o(0xd3); /* shl/shr/sar %cl, r */
 858                         o(0xc0 + r + opc*8);
 859         }
 860         vtop--;
 861         break;
 862     case TOK_UMOD:
 863                 opc = 4;
 864         uu = 1;
 865         goto divmod;
 866     case TOK_UDIV:
 867     case TOK_UMULL:
 868                 opc = 6;
 869         uu = 1;
 870         goto divmod;
 871     case '/':
 872     case '%':
 873     case TOK_PDIV:
 874                 opc = 7;
 875         uu = 0;
 876         divmod:
 877         /* first operand must be in eax */
 878         /* XXX: need better constraint for second operand */
 879                 if(!tt2){
 880                         gv2(RC_EAX, RC_INT2);
 881                         fr = vtop[0].r;
 882                 }else{
 883                         vswap();
 884                         gv(RC_EAX);
 885                         vswap();
 886                 }
 887                 save_reg(TREG_EDX);
 888                 if (op == TOK_UMULL) {
 889             o(0xf7); /* mul fr */
 890                         vtop->r2 = TREG_EDX;
 891                 }else{
 892                         o(uu ? 0xd231 : 0x99); /* xor %edx,%edx : cdq RDX:RAX <- sign-extend of RAX. */
 893                         o(0xf7); /* div fr, %eax */
 894                 }
 895                 if(fr >= VT_CONST)
 896                         gen_modrm(opc, fr, vtop->sym, fc);
 897                 else
 898                         o(0xc0 + fr + opc*8);
 899         if (op == '%' || op == TOK_UMOD)
 900             r = TREG_EDX;
 901         else
 902             r = TREG_EAX;
 903                 vtop--;
 904         vtop->r = r;
 905         break;
 906     default:
 907         opc = 7;
 908         goto gen_op8;
 909     }
 910 }
 911
 912 /* generate a floating point operation 'v = t1 op t2' instruction. The
 913    two operands are guaranted to have the same floating point type */
 914 /* XXX: need to use ST1 too */
 915 ST_FUNC void gen_opf(int op)
 916 {
 917     int a, ft, fc, swapped, r;
 918
 919     /* convert constants to memory references */
 920     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 921         vswap();
 922         gv(RC_FLOAT);
 923         vswap();
 924     }
 925     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 926         gv(RC_FLOAT);
 927
 928     /* must put at least one value in the floating point register */
 929     if ((vtop[-1].r & VT_LVAL) &&
 930         (vtop[0].r & VT_LVAL)) {
 931         vswap();
 932         gv(RC_FLOAT);
 933         vswap();
 934     }
 935     swapped = 0;
 936     /* swap the stack if needed so that t1 is the register and t2 is
 937        the memory reference */
 938     if (vtop[-1].r & VT_LVAL) {
 939         vswap();
 940         swapped = 1;
 941     }
 942     if (op >= TOK_ULT && op <= TOK_GT) {
 943         /* load on stack second operand */
 944         load(TREG_ST0, vtop);
 945         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 946         if (op == TOK_GE || op == TOK_GT)
 947             swapped = !swapped;
 948         else if (op == TOK_EQ || op == TOK_NE)
 949             swapped = 0;
 950         if (swapped)
 951             o(0xc9d9); /* fxch %st(1) */
 952         if (op == TOK_EQ || op == TOK_NE)
 953             o(0xe9da); /* fucompp */
 954         else
 955             o(0xd9de); /* fcompp */
 956         o(0xe0df); /* fnstsw %ax */
 957         if (op == TOK_EQ) {
 958             o(0x45e480); /* and $0x45, %ah */
 959             o(0x40fC80); /* cmp $0x40, %ah */
 960         } else if (op == TOK_NE) {
 961             o(0x45e480); /* and $0x45, %ah */
 962             o(0x40f480); /* xor $0x40, %ah */
 963             op = TOK_NE;
 964         } else if (op == TOK_GE || op == TOK_LE) {
 965             o(0x05c4f6); /* test $0x05, %ah */
 966             op = TOK_EQ;
 967         } else {
 968             o(0x45c4f6); /* test $0x45, %ah */
 969             op = TOK_EQ;
 970         }
 971         vtop--;
 972         vtop->r = VT_CMP;
 973         vtop->c.i = op;
 974     } else {
 975         /* no memory reference possible for long double operations */
 976         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 977             load(TREG_ST0, vtop);
 978             swapped = !swapped;
 979         }
 980
 981         switch(op) {
 982         default:
 983         case '+':
 984             a = 0;
 985             break;
 986         case '-':
 987             a = 4;
 988             if (swapped)
 989                 a++;
 990             break;
 991         case '*':
 992             a = 1;
 993             break;
 994         case '/':
 995             a = 6;
 996             if (swapped)
 997                 a++;
 998             break;
 999         }
1000         ft = vtop->type.t;
1001         fc = vtop->c.ul;
1002         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
1003             o(0xde); /* fxxxp %st, %st(1) */
1004             o(0xc1 + (a << 3));
1005         } else {
1006             /* if saved lvalue, then we must reload it */
1007             r = vtop->r;
1008             if ((r & VT_VALMASK) == VT_LLOCAL) {
1009                 SValue v1;
1010                 r = get_reg(RC_INT);
1011                 v1.type.t = VT_INT;
1012                 v1.r = VT_LOCAL | VT_LVAL;
1013                 v1.c.ul = fc;
1014                 load(r, &v1);
1015                 fc = 0;
1016             }
1017
1018             if ((ft & VT_BTYPE) == VT_DOUBLE)
1019                 o(0xdc);
1020             else
1021                 o(0xd8);
1022             gen_modrm(a, r, vtop->sym, fc);
1023         }
1024         vtop--;
1025     }
1026 }
1027
1028 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1029    and 'long long' cases. */
1030 ST_FUNC void gen_cvt_itof(int t)
1031 {
1032     save_reg(TREG_ST0);
1033     gv(RC_INT);
1034     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
1035         /* signed long long to float/double/long double (unsigned case
1036            is handled generically) */
1037         o(0x50 + vtop->r2); /* push r2 */
1038         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1039         o(0x242cdf); /* fildll (%esp) */
1040         o(0x08c483); /* add $8, %esp */
1041     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
1042                (VT_INT | VT_UNSIGNED)) {
1043         /* unsigned int to float/double/long double */
1044         o(0x6a); /* push $0 */
1045         g(0x00);
1046         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1047         o(0x242cdf); /* fildll (%esp) */
1048         o(0x08c483); /* add $8, %esp */
1049     } else {
1050         /* int to float/double/long double */
1051         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1052         o(0x2404db); /* fildl (%esp) */
1053         o(0x04c483); /* add $4, %esp */
1054     }
1055     vtop->r = TREG_ST0;
1056 }
1057
1058 /* convert fp to int 't' type */
1059 ST_FUNC void gen_cvt_ftoi(int t)
1060 {
1061     int bt = vtop->type.t & VT_BTYPE;
1062     if (bt == VT_FLOAT)
1063         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1064     else if (bt == VT_LDOUBLE)
1065         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1066     else
1067         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1068     vswap();
1069     gfunc_call(1);
1070     vpushi(0);
1071     vtop->r = REG_IRET;
1072     vtop->r2 = REG_LRET;
1073 }
1074
1075 /* convert from one floating point type to another */
1076 ST_FUNC void gen_cvt_ftof(int t)
1077 {
1078     /* all we have to do on i386 is to put the float in a register */
1079     gv(RC_FLOAT);
1080 }
1081
1082 /* computed goto support */
1083 ST_FUNC void ggoto(void)
1084 {
1085     gcall_or_jmp(1);
1086     vtop--;
1087 }
1088
1089 /* bound check support functions */
1090 #ifdef CONFIG_TCC_BCHECK
1091
1092 /* generate a bounded pointer addition */
1093 ST_FUNC void gen_bounded_ptr_add(void)
1094 {
1095     /* prepare fast i386 function call (args in eax and edx) */
1096     gv2(RC_EAX, RC_EDX);
1097     /* save all temporary registers */
1098     vtop -= 2;
1099     save_regs(0);
1100     /* do a fast function call */
1101     gen_static_call(TOK___bound_ptr_add);
1102     /* returned pointer is in eax */
1103     vtop++;
1104     vtop->r = TREG_EAX | VT_BOUNDED;
1105     /* address of bounding function call point */
1106     vtop->c.ul = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1107 }
1108
1109 /* patch pointer addition in vtop so that pointer dereferencing is
1110    also tested */
1111 ST_FUNC void gen_bounded_ptr_deref(void)
1112 {
1113     int func;
1114     int size, align;
1115     Elf32_Rel *rel;
1116     Sym *sym;
1117
1118     size = 0;
1119     /* XXX: put that code in generic part of tcc */
1120     if (!is_float(vtop->type.t)) {
1121         if (vtop->r & VT_LVAL_BYTE)
1122             size = 1;
1123         else if (vtop->r & VT_LVAL_SHORT)
1124             size = 2;
1125     }
1126     if (!size)
1127         size = type_size(&vtop->type, &align);
1128     switch(size) {
1129     case  1: func = TOK___bound_ptr_indir1; break;
1130     case  2: func = TOK___bound_ptr_indir2; break;
1131     case  4: func = TOK___bound_ptr_indir4; break;
1132     case  8: func = TOK___bound_ptr_indir8; break;
1133     case 12: func = TOK___bound_ptr_indir12; break;
1134     case 16: func = TOK___bound_ptr_indir16; break;
1135     default:
1136         tcc_error("unhandled size when dereferencing bounded pointer");
1137         func = 0;
1138         break;
1139     }
1140
1141     /* patch relocation */
1142     /* XXX: find a better solution ? */
1143     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.ul);
1144     sym = external_global_sym(func, &func_old_type, 0);
1145     if (!sym->c)
1146         put_extern_sym(sym, NULL, 0, 0);
1147     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1148 }
1149 #endif
1150
1151 /* Save the stack pointer onto the stack */
1152 ST_FUNC void gen_vla_sp_save(int addr) {
1153     /* mov %esp,addr(%ebp)*/
1154     o(0x89);
1155     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1156 }
1157
1158 /* Restore the SP from a location on the stack */
1159 ST_FUNC void gen_vla_sp_restore(int addr) {
1160     o(0x8b);
1161     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1162 }
1163
1164 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1165 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1166 #ifdef TCC_TARGET_PE
1167     /* alloca does more than just adjust %rsp on Windows */
1168     vpush_global_sym(&func_old_type, TOK_alloca);
1169     vswap(); /* Move alloca ref past allocation size */
1170     gfunc_call(1);
1171     vset(type, REG_IRET, 0);
1172 #else
1173     int r;
1174     r = gv(RC_INT); /* allocation size */
1175     /* sub r,%rsp */
1176     o(0x2b);
1177     o(0xe0 | r);
1178     /* We align to 16 bytes rather than align */
1179     /* and ~15, %esp */
1180     o(0xf0e483);
1181     /* mov %esp, r */
1182     o(0x89);
1183     o(0xe0 | r);
1184     vpop();
1185     vset(type, r, 0);
1186 #endif
1187 }
1188
1189 /* end of X86 code generator */
1190 /*************************************************************/
1191 #endif
1192 /*************************************************************/