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         5
  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
  38 #define RC_IRET    RC_EAX /* function return: integer register */
  39 #define RC_LRET    RC_EDX /* function return: second integer register */
  40 #define RC_FRET    RC_ST0 /* function return: float register */
  41
  42 /* pretty names for the registers */
  43 enum {
  44     TREG_EAX = 0,
  45     TREG_ECX,
  46     TREG_EDX,
  47     TREG_EBX,
  48     TREG_ST0,
  49     TREG_ESP = 4
  50 };
  51
  52 /* return registers for function */
  53 #define REG_IRET TREG_EAX /* single word int return register */
  54 #define REG_LRET TREG_EDX /* second word return register (for long long) */
  55 #define REG_FRET TREG_ST0 /* float return register */
  56
  57 /* defined if function parameters must be evaluated in reverse order */
  58 #define INVERT_FUNC_PARAMS
  59
  60 /* defined if structures are passed as pointers. Otherwise structures
  61    are directly pushed on stack. */
  62 /* #define FUNC_STRUCT_PARAM_AS_PTR */
  63
  64 /* pointer size, in bytes */
  65 #define PTR_SIZE 4
  66
  67 /* long double size and alignment, in bytes */
  68 #define LDOUBLE_SIZE  12
  69 #define LDOUBLE_ALIGN 4
  70 /* maximum alignment (for aligned attribute support) */
  71 #define MAX_ALIGN     8
  72
  73
  74 #define psym oad
  75
  76 /******************************************************/
  77 /* ELF defines */
  78
  79 #define EM_TCC_TARGET EM_386
  80
  81 /* relocation type for 32 bit data relocation */
  82 #define R_DATA_32   R_386_32
  83 #define R_DATA_PTR  R_386_32
  84 #define R_JMP_SLOT  R_386_JMP_SLOT
  85 #define R_GLOB_DAT  R_386_GLOB_DAT
  86 #define R_COPY      R_386_COPY
  87
  88 #define ELF_START_ADDR 0x08048000
  89 #define ELF_PAGE_SIZE  0x1000
  90
  91 /******************************************************/
  92 #else /* ! TARGET_DEFS_ONLY */
  93 /******************************************************/
  94 #include "tcc.h"
  95
  96 /* define to 1/0 to [not] have EBX as 4th register */
  97 #define USE_EBX 1
  98
  99 ST_DATA const int reg_classes[NB_REGS] = {
 100     /* eax */ RC_INT | RC_EAX,
 101     /* ecx */ RC_INT | RC_ECX,
 102     /* edx */ RC_INT | RC_EDX,
 103     /* ebx */ (RC_INT | RC_EBX) * USE_EBX,
 104     /* st0 */ RC_FLOAT | RC_ST0,
 105 };
 106
 107 static unsigned long func_sub_sp_offset;
 108 static int func_ret_sub;
 109 #ifdef CONFIG_TCC_BCHECK
 110 static addr_t func_bound_offset;
 111 static unsigned long func_bound_ind;
 112 #endif
 113
 114 /* XXX: make it faster ? */
 115 ST_FUNC void g(int c)
 116 {
 117     int ind1;
 118     ind1 = ind + 1;
 119     if (ind1 > cur_text_section->data_allocated)
 120         section_realloc(cur_text_section, ind1);
 121     cur_text_section->data[ind] = c;
 122     ind = ind1;
 123 }
 124
 125 ST_FUNC void o(unsigned int c)
 126 {
 127     while (c) {
 128         g(c);
 129         c = c >> 8;
 130     }
 131 }
 132
 133 ST_FUNC void gen_le16(int v)
 134 {
 135     g(v);
 136     g(v >> 8);
 137 }
 138
 139 ST_FUNC void gen_le32(int c)
 140 {
 141     g(c);
 142     g(c >> 8);
 143     g(c >> 16);
 144     g(c >> 24);
 145 }
 146
 147 /* output a symbol and patch all calls to it */
 148 ST_FUNC void gsym_addr(int t, int a)
 149 {
 150     while (t) {
 151         unsigned char *ptr = cur_text_section->data + t;
 152         uint32_t n = read32le(ptr); /* next value */
 153         write32le(ptr, a - t - 4);
 154         t = n;
 155     }
 156 }
 157
 158 ST_FUNC void gsym(int t)
 159 {
 160     gsym_addr(t, ind);
 161 }
 162
 163 /* psym is used to put an instruction with a data field which is a
 164    reference to a symbol. It is in fact the same as oad ! */
 165 #define psym oad
 166
 167 /* instruction + 4 bytes data. Return the address of the data */
 168 ST_FUNC int oad(int c, int s)
 169 {
 170     int ind1;
 171
 172     o(c);
 173     ind1 = ind + 4;
 174     if (ind1 > cur_text_section->data_allocated)
 175         section_realloc(cur_text_section, ind1);
 176     write32le(cur_text_section->data + ind, s);
 177     s = ind;
 178     ind = ind1;
 179     return s;
 180 }
 181
 182 /* output constant with relocation if 'r & VT_SYM' is true */
 183 ST_FUNC void gen_addr32(int r, Sym *sym, int c)
 184 {
 185     if (r & VT_SYM)
 186         greloc(cur_text_section, sym, ind, R_386_32);
 187     gen_le32(c);
 188 }
 189
 190 ST_FUNC void gen_addrpc32(int r, Sym *sym, int c)
 191 {
 192     if (r & VT_SYM)
 193         greloc(cur_text_section, sym, ind, R_386_PC32);
 194     gen_le32(c - 4);
 195 }
 196
 197 /* generate a modrm reference. 'op_reg' contains the addtionnal 3
 198    opcode bits */
 199 static void gen_modrm(int op_reg, int r, Sym *sym, int c)
 200 {
 201     op_reg = op_reg << 3;
 202     if ((r & VT_VALMASK) == VT_CONST) {
 203         /* constant memory reference */
 204         o(0x05 | op_reg);
 205         gen_addr32(r, sym, c);
 206     } else if ((r & VT_VALMASK) == VT_LOCAL) {
 207         /* currently, we use only ebp as base */
 208         if (c == (char)c) {
 209             /* short reference */
 210             o(0x45 | op_reg);
 211             g(c);
 212         } else {
 213             oad(0x85 | op_reg, c);
 214         }
 215     } else {
 216         g(0x00 | op_reg | (r & VT_VALMASK));
 217     }
 218 }
 219
 220 /* load 'r' from value 'sv' */
 221 ST_FUNC void load(int r, SValue *sv)
 222 {
 223     int v, t, ft, fc, fr;
 224     SValue v1;
 225
 226 #ifdef TCC_TARGET_PE
 227     SValue v2;
 228     sv = pe_getimport(sv, &v2);
 229 #endif
 230
 231     fr = sv->r;
 232     ft = sv->type.t;
 233     fc = sv->c.i;
 234
 235     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 236
 237     v = fr & VT_VALMASK;
 238     if (fr & VT_LVAL) {
 239         if (v == VT_LLOCAL) {
 240             v1.type.t = VT_INT;
 241             v1.r = VT_LOCAL | VT_LVAL;
 242             v1.c.i = fc;
 243             fr = r;
 244             if (!(reg_classes[fr] & RC_INT))
 245                 fr = get_reg(RC_INT);
 246             load(fr, &v1);
 247         }
 248         if ((ft & VT_BTYPE) == VT_FLOAT) {
 249             o(0xd9); /* flds */
 250             r = 0;
 251         } else if ((ft & VT_BTYPE) == VT_DOUBLE) {
 252             o(0xdd); /* fldl */
 253             r = 0;
 254         } else if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 255             o(0xdb); /* fldt */
 256             r = 5;
 257         } else if ((ft & VT_TYPE) == VT_BYTE || (ft & VT_TYPE) == VT_BOOL) {
 258             o(0xbe0f);   /* movsbl */
 259         } else if ((ft & VT_TYPE) == (VT_BYTE | VT_UNSIGNED)) {
 260             o(0xb60f);   /* movzbl */
 261         } else if ((ft & VT_TYPE) == VT_SHORT) {
 262             o(0xbf0f);   /* movswl */
 263         } else if ((ft & VT_TYPE) == (VT_SHORT | VT_UNSIGNED)) {
 264             o(0xb70f);   /* movzwl */
 265         } else {
 266             o(0x8b);     /* movl */
 267         }
 268         gen_modrm(r, fr, sv->sym, fc);
 269     } else {
 270         if (v == VT_CONST) {
 271             o(0xb8 + r); /* mov $xx, r */
 272             gen_addr32(fr, sv->sym, fc);
 273         } else if (v == VT_LOCAL) {
 274             if (fc) {
 275                 o(0x8d); /* lea xxx(%ebp), r */
 276                 gen_modrm(r, VT_LOCAL, sv->sym, fc);
 277             } else {
 278                 o(0x89);
 279                 o(0xe8 + r); /* mov %ebp, r */
 280             }
 281         } else if (v == VT_CMP) {
 282             oad(0xb8 + r, 0); /* mov $0, r */
 283             o(0x0f); /* setxx %br */
 284             o(fc);
 285             o(0xc0 + r);
 286         } else if (v == VT_JMP || v == VT_JMPI) {
 287             t = v & 1;
 288             oad(0xb8 + r, t); /* mov $1, r */
 289             o(0x05eb); /* jmp after */
 290             gsym(fc);
 291             oad(0xb8 + r, t ^ 1); /* mov $0, r */
 292         } else if (v != r) {
 293             o(0x89);
 294             o(0xc0 + r + v * 8); /* mov v, r */
 295         }
 296     }
 297 }
 298
 299 /* store register 'r' in lvalue 'v' */
 300 ST_FUNC void store(int r, SValue *v)
 301 {
 302     int fr, bt, ft, fc;
 303
 304 #ifdef TCC_TARGET_PE
 305     SValue v2;
 306     v = pe_getimport(v, &v2);
 307 #endif
 308
 309     ft = v->type.t;
 310     fc = v->c.i;
 311     fr = v->r & VT_VALMASK;
 312     ft &= ~(VT_VOLATILE | VT_CONSTANT);
 313     bt = ft & VT_BTYPE;
 314     /* XXX: incorrect if float reg to reg */
 315     if (bt == VT_FLOAT) {
 316         o(0xd9); /* fsts */
 317         r = 2;
 318     } else if (bt == VT_DOUBLE) {
 319         o(0xdd); /* fstpl */
 320         r = 2;
 321     } else if (bt == VT_LDOUBLE) {
 322         o(0xc0d9); /* fld %st(0) */
 323         o(0xdb); /* fstpt */
 324         r = 7;
 325     } else {
 326         if (bt == VT_SHORT)
 327             o(0x66);
 328         if (bt == VT_BYTE || bt == VT_BOOL)
 329             o(0x88);
 330         else
 331             o(0x89);
 332     }
 333     if (fr == VT_CONST ||
 334         fr == VT_LOCAL ||
 335         (v->r & VT_LVAL)) {
 336         gen_modrm(r, v->r, v->sym, fc);
 337     } else if (fr != r) {
 338         o(0xc0 + fr + r * 8); /* mov r, fr */
 339     }
 340 }
 341
 342 static void gadd_sp(int val)
 343 {
 344     if (val == (char)val) {
 345         o(0xc483);
 346         g(val);
 347     } else {
 348         oad(0xc481, val); /* add $xxx, %esp */
 349     }
 350 }
 351
 352 static void gen_static_call(int v)
 353 {
 354     Sym *sym;
 355
 356     sym = external_global_sym(v, &func_old_type, 0);
 357     oad(0xe8, -4);
 358     greloc(cur_text_section, sym, ind-4, R_386_PC32);
 359 }
 360
 361 /* 'is_jmp' is '1' if it is a jump */
 362 static void gcall_or_jmp(int is_jmp)
 363 {
 364     int r;
 365     if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 366         int rt;
 367         /* constant case */
 368         if (vtop->r & VT_SYM) {
 369             /* relocation case */
 370             greloc(cur_text_section, vtop->sym,
 371                    ind + 1, R_386_PC32);
 372         } else {
 373             /* put an empty PC32 relocation */
 374             put_elf_reloc(symtab_section, cur_text_section,
 375                           ind + 1, R_386_PC32, 0);
 376         }
 377         oad(0xe8 + is_jmp, vtop->c.i - 4); /* call/jmp im */
 378         /* extend the return value to the whole register if necessary
 379            visual studio and gcc do not always set the whole eax register
 380            when assigning the return value of a function  */
 381         rt = vtop->type.ref->type.t;
 382         switch (rt & VT_BTYPE) {
 383             case VT_BYTE:
 384                 if (rt & VT_UNSIGNED) {
 385                     o(0xc0b60f); /* movzx %al, %eax */
 386                 }
 387                 else {
 388                     o(0xc0be0f); /* movsx %al, %eax */
 389                 }
 390                 break;
 391             case VT_SHORT:
 392                 if (rt & VT_UNSIGNED) {
 393                     o(0xc0b70f); /* movzx %ax, %eax */
 394                 }
 395                 else {
 396                     o(0xc0bf0f); /* movsx %ax, %eax */
 397                 }
 398                 break;
 399             default:
 400                 break;
 401         }
 402     } else {
 403         /* otherwise, indirect call */
 404         r = gv(RC_INT);
 405         o(0xff); /* call/jmp *r */
 406         o(0xd0 + r + (is_jmp << 4));
 407     }
 408 }
 409
 410 static uint8_t fastcall_regs[3] = { TREG_EAX, TREG_EDX, TREG_ECX };
 411 static uint8_t fastcallw_regs[2] = { TREG_ECX, TREG_EDX };
 412
 413 /* Return the number of registers needed to return the struct, or 0 if
 414    returning via struct pointer. */
 415 ST_FUNC int gfunc_sret(CType *vt, int variadic, CType *ret, int *ret_align, int *regsize)
 416 {
 417 #ifdef TCC_TARGET_PE
 418     int size, align;
 419
 420     *ret_align = 1; // Never have to re-align return values for x86
 421     *regsize = 4;
 422     size = type_size(vt, &align);
 423     if (size > 8) {
 424         return 0;
 425     } else if (size > 4) {
 426         ret->ref = NULL;
 427         ret->t = VT_LLONG;
 428         return 1;
 429     } else {
 430         ret->ref = NULL;
 431         ret->t = VT_INT;
 432         return 1;
 433     }
 434 #else
 435     *ret_align = 1; // Never have to re-align return values for x86
 436     return 0;
 437 #endif
 438 }
 439
 440 /* Generate function call. The function address is pushed first, then
 441    all the parameters in call order. This functions pops all the
 442    parameters and the function address. */
 443 ST_FUNC void gfunc_call(int nb_args)
 444 {
 445     int size, align, r, args_size, i, func_call;
 446     Sym *func_sym;
 447
 448     args_size = 0;
 449     for(i = 0;i < nb_args; i++) {
 450         if ((vtop->type.t & VT_BTYPE) == VT_STRUCT) {
 451             size = type_size(&vtop->type, &align);
 452             /* align to stack align size */
 453             size = (size + 3) & ~3;
 454             /* allocate the necessary size on stack */
 455             oad(0xec81, size); /* sub $xxx, %esp */
 456             /* generate structure store */
 457             r = get_reg(RC_INT);
 458             o(0x89); /* mov %esp, r */
 459             o(0xe0 + r);
 460             vset(&vtop->type, r | VT_LVAL, 0);
 461             vswap();
 462             vstore();
 463             args_size += size;
 464         } else if (is_float(vtop->type.t)) {
 465             gv(RC_FLOAT); /* only one float register */
 466             if ((vtop->type.t & VT_BTYPE) == VT_FLOAT)
 467                 size = 4;
 468             else if ((vtop->type.t & VT_BTYPE) == VT_DOUBLE)
 469                 size = 8;
 470             else
 471                 size = 12;
 472             oad(0xec81, size); /* sub $xxx, %esp */
 473             if (size == 12)
 474                 o(0x7cdb);
 475             else
 476                 o(0x5cd9 + size - 4); /* fstp[s|l] 0(%esp) */
 477             g(0x24);
 478             g(0x00);
 479             args_size += size;
 480         } else {
 481             /* simple type (currently always same size) */
 482             /* XXX: implicit cast ? */
 483             r = gv(RC_INT);
 484             if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
 485                 size = 8;
 486                 o(0x50 + vtop->r2); /* push r */
 487             } else {
 488                 size = 4;
 489             }
 490             o(0x50 + r); /* push r */
 491             args_size += size;
 492         }
 493         vtop--;
 494     }
 495     save_regs(0); /* save used temporary registers */
 496     func_sym = vtop->type.ref;
 497     func_call = func_sym->a.func_call;
 498     /* fast call case */
 499     if ((func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) ||
 500         func_call == FUNC_FASTCALLW) {
 501         int fastcall_nb_regs;
 502         uint8_t *fastcall_regs_ptr;
 503         if (func_call == FUNC_FASTCALLW) {
 504             fastcall_regs_ptr = fastcallw_regs;
 505             fastcall_nb_regs = 2;
 506         } else {
 507             fastcall_regs_ptr = fastcall_regs;
 508             fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 509         }
 510         for(i = 0;i < fastcall_nb_regs; i++) {
 511             if (args_size <= 0)
 512                 break;
 513             o(0x58 + fastcall_regs_ptr[i]); /* pop r */
 514             /* XXX: incorrect for struct/floats */
 515             args_size -= 4;
 516         }
 517     }
 518 #ifndef TCC_TARGET_PE
 519     else if ((vtop->type.ref->type.t & VT_BTYPE) == VT_STRUCT)
 520         args_size -= 4;
 521 #endif
 522     gcall_or_jmp(0);
 523
 524     if (args_size && func_call != FUNC_STDCALL)
 525         gadd_sp(args_size);
 526     vtop--;
 527 }
 528
 529 #ifdef TCC_TARGET_PE
 530 #define FUNC_PROLOG_SIZE (10 + USE_EBX)
 531 #else
 532 #define FUNC_PROLOG_SIZE (9 + USE_EBX)
 533 #endif
 534
 535 /* generate function prolog of type 't' */
 536 ST_FUNC void gfunc_prolog(CType *func_type)
 537 {
 538     int addr, align, size, func_call, fastcall_nb_regs;
 539     int param_index, param_addr;
 540     uint8_t *fastcall_regs_ptr;
 541     Sym *sym;
 542     CType *type;
 543
 544     sym = func_type->ref;
 545     func_call = sym->a.func_call;
 546     addr = 8;
 547     loc = 0;
 548     func_vc = 0;
 549
 550     if (func_call >= FUNC_FASTCALL1 && func_call <= FUNC_FASTCALL3) {
 551         fastcall_nb_regs = func_call - FUNC_FASTCALL1 + 1;
 552         fastcall_regs_ptr = fastcall_regs;
 553     } else if (func_call == FUNC_FASTCALLW) {
 554         fastcall_nb_regs = 2;
 555         fastcall_regs_ptr = fastcallw_regs;
 556     } else {
 557         fastcall_nb_regs = 0;
 558         fastcall_regs_ptr = NULL;
 559     }
 560     param_index = 0;
 561
 562     ind += FUNC_PROLOG_SIZE;
 563     func_sub_sp_offset = ind;
 564     /* if the function returns a structure, then add an
 565        implicit pointer parameter */
 566     func_vt = sym->type;
 567     func_var = (sym->c == FUNC_ELLIPSIS);
 568 #ifdef TCC_TARGET_PE
 569     size = type_size(&func_vt,&align);
 570     if (((func_vt.t & VT_BTYPE) == VT_STRUCT) && (size > 8)) {
 571 #else
 572     if ((func_vt.t & VT_BTYPE) == VT_STRUCT) {
 573 #endif
 574         /* XXX: fastcall case ? */
 575         func_vc = addr;
 576         addr += 4;
 577         param_index++;
 578     }
 579     /* define parameters */
 580     while ((sym = sym->next) != NULL) {
 581         type = &sym->type;
 582         size = type_size(type, &align);
 583         size = (size + 3) & ~3;
 584 #ifdef FUNC_STRUCT_PARAM_AS_PTR
 585         /* structs are passed as pointer */
 586         if ((type->t & VT_BTYPE) == VT_STRUCT) {
 587             size = 4;
 588         }
 589 #endif
 590         if (param_index < fastcall_nb_regs) {
 591             /* save FASTCALL register */
 592             loc -= 4;
 593             o(0x89);     /* movl */
 594             gen_modrm(fastcall_regs_ptr[param_index], VT_LOCAL, NULL, loc);
 595             param_addr = loc;
 596         } else {
 597             param_addr = addr;
 598             addr += size;
 599         }
 600         sym_push(sym->v & ~SYM_FIELD, type,
 601                  VT_LOCAL | lvalue_type(type->t), param_addr);
 602         param_index++;
 603     }
 604     func_ret_sub = 0;
 605     /* pascal type call ? */
 606     if (func_call == FUNC_STDCALL)
 607         func_ret_sub = addr - 8;
 608 #ifndef TCC_TARGET_PE
 609     else if (func_vc)
 610         func_ret_sub = 4;
 611 #endif
 612
 613 #ifdef CONFIG_TCC_BCHECK
 614     /* leave some room for bound checking code */
 615     if (tcc_state->do_bounds_check) {
 616         func_bound_offset = lbounds_section->data_offset;
 617         func_bound_ind = ind;
 618         oad(0xb8, 0); /* lbound section pointer */
 619         oad(0xb8, 0); /* call to function */
 620     }
 621 #endif
 622 }
 623
 624 /* generate function epilog */
 625 ST_FUNC void gfunc_epilog(void)
 626 {
 627     addr_t v, saved_ind;
 628
 629 #ifdef CONFIG_TCC_BCHECK
 630     if (tcc_state->do_bounds_check
 631      && func_bound_offset != lbounds_section->data_offset) {
 632         addr_t saved_ind;
 633         addr_t *bounds_ptr;
 634         Sym *sym_data;
 635
 636         /* add end of table info */
 637         bounds_ptr = section_ptr_add(lbounds_section, sizeof(addr_t));
 638         *bounds_ptr = 0;
 639
 640         /* generate bound local allocation */
 641         saved_ind = ind;
 642         ind = func_bound_ind;
 643         sym_data = get_sym_ref(&char_pointer_type, lbounds_section,
 644                                func_bound_offset, lbounds_section->data_offset);
 645         greloc(cur_text_section, sym_data,
 646                ind + 1, R_386_32);
 647         oad(0xb8, 0); /* mov %eax, xxx */
 648         gen_static_call(TOK___bound_local_new);
 649         ind = saved_ind;
 650
 651         /* generate bound check local freeing */
 652         o(0x5250); /* save returned value, if any */
 653         greloc(cur_text_section, sym_data, ind + 1, R_386_32);
 654         oad(0xb8, 0); /* mov %eax, xxx */
 655         gen_static_call(TOK___bound_local_delete);
 656         o(0x585a); /* restore returned value, if any */
 657     }
 658 #endif
 659     o(0x5b * USE_EBX); /* pop ebx */
 660     o(0xc9); /* leave */
 661     if (func_ret_sub == 0) {
 662         o(0xc3); /* ret */
 663     } else {
 664         o(0xc2); /* ret n */
 665         g(func_ret_sub);
 666         g(func_ret_sub >> 8);
 667     }
 668     /* align local size to word & save local variables */
 669
 670     v = (-loc + 3) & -4;
 671     saved_ind = ind;
 672     ind = func_sub_sp_offset - FUNC_PROLOG_SIZE;
 673 #ifdef TCC_TARGET_PE
 674     if (v >= 4096) {
 675         oad(0xb8, v); /* mov stacksize, %eax */
 676         gen_static_call(TOK___chkstk); /* call __chkstk, (does the stackframe too) */
 677     } else
 678 #endif
 679     {
 680         o(0xe58955);  /* push %ebp, mov %esp, %ebp */
 681         o(0xec81);  /* sub esp, stacksize */
 682         gen_le32(v);
 683 #ifdef TCC_TARGET_PE
 684         o(0x90);  /* adjust to FUNC_PROLOG_SIZE */
 685 #endif
 686     }
 687     o(0x53 * USE_EBX); /* push ebx */
 688     ind = saved_ind;
 689 }
 690
 691 /* generate a jump to a label */
 692 ST_FUNC int gjmp(int t)
 693 {
 694     return psym(0xe9, t);
 695 }
 696
 697 /* generate a jump to a fixed address */
 698 ST_FUNC void gjmp_addr(int a)
 699 {
 700     int r;
 701     r = a - ind - 2;
 702     if (r == (char)r) {
 703         g(0xeb);
 704         g(r);
 705     } else {
 706         oad(0xe9, a - ind - 5);
 707     }
 708 }
 709
 710 ST_FUNC void gtst_addr(int inv, int a)
 711 {
 712     inv ^= (vtop--)->c.i;
 713     a -= ind + 2;
 714     if (a == (char)a) {
 715         g(inv - 32);
 716         g(a);
 717     } else {
 718         g(0x0f);
 719         oad(inv - 16, a - 4);
 720     }
 721 }
 722
 723 /* generate a test. set 'inv' to invert test. Stack entry is popped */
 724 ST_FUNC int gtst(int inv, int t)
 725 {
 726     int v = vtop->r & VT_VALMASK;
 727     if (v == VT_CMP) {
 728         /* fast case : can jump directly since flags are set */
 729         g(0x0f);
 730         t = psym((vtop->c.i - 16) ^ inv, t);
 731     } else if (v == VT_JMP || v == VT_JMPI) {
 732         /* && or || optimization */
 733         if ((v & 1) == inv) {
 734             /* insert vtop->c jump list in t */
 735             uint32_t n1, n = vtop->c.i;
 736             if (n) {
 737                 while ((n1 = read32le(cur_text_section->data + n)))
 738                     n = n1;
 739                 write32le(cur_text_section->data + n, t);
 740                 t = vtop->c.i;
 741             }
 742         } else {
 743             t = gjmp(t);
 744             gsym(vtop->c.i);
 745         }
 746     }
 747     vtop--;
 748     return t;
 749 }
 750
 751 /* generate an integer binary operation */
 752 ST_FUNC void gen_opi(int op)
 753 {
 754     int r, fr, opc, c;
 755
 756     switch(op) {
 757     case '+':
 758     case TOK_ADDC1: /* add with carry generation */
 759         opc = 0;
 760     gen_op8:
 761         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 762             /* constant case */
 763             vswap();
 764             r = gv(RC_INT);
 765             vswap();
 766             c = vtop->c.i;
 767             if (c == (char)c) {
 768                 /* generate inc and dec for smaller code */
 769                 if (c==1 && opc==0 && op != TOK_ADDC1) {
 770                     o (0x40 | r); // inc
 771                 } else if (c==1 && opc==5 && op != TOK_SUBC1) {
 772                     o (0x48 | r); // dec
 773                 } else {
 774                     o(0x83);
 775                     o(0xc0 | (opc << 3) | r);
 776                     g(c);
 777                 }
 778             } else {
 779                 o(0x81);
 780                 oad(0xc0 | (opc << 3) | r, c);
 781             }
 782         } else {
 783             gv2(RC_INT, RC_INT);
 784             r = vtop[-1].r;
 785             fr = vtop[0].r;
 786             o((opc << 3) | 0x01);
 787             o(0xc0 + r + fr * 8);
 788         }
 789         vtop--;
 790         if (op >= TOK_ULT && op <= TOK_GT) {
 791             vtop->r = VT_CMP;
 792             vtop->c.i = op;
 793         }
 794         break;
 795     case '-':
 796     case TOK_SUBC1: /* sub with carry generation */
 797         opc = 5;
 798         goto gen_op8;
 799     case TOK_ADDC2: /* add with carry use */
 800         opc = 2;
 801         goto gen_op8;
 802     case TOK_SUBC2: /* sub with carry use */
 803         opc = 3;
 804         goto gen_op8;
 805     case '&':
 806         opc = 4;
 807         goto gen_op8;
 808     case '^':
 809         opc = 6;
 810         goto gen_op8;
 811     case '|':
 812         opc = 1;
 813         goto gen_op8;
 814     case '*':
 815         gv2(RC_INT, RC_INT);
 816         r = vtop[-1].r;
 817         fr = vtop[0].r;
 818         vtop--;
 819         o(0xaf0f); /* imul fr, r */
 820         o(0xc0 + fr + r * 8);
 821         break;
 822     case TOK_SHL:
 823         opc = 4;
 824         goto gen_shift;
 825     case TOK_SHR:
 826         opc = 5;
 827         goto gen_shift;
 828     case TOK_SAR:
 829         opc = 7;
 830     gen_shift:
 831         opc = 0xc0 | (opc << 3);
 832         if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == VT_CONST) {
 833             /* constant case */
 834             vswap();
 835             r = gv(RC_INT);
 836             vswap();
 837             c = vtop->c.i & 0x1f;
 838             o(0xc1); /* shl/shr/sar $xxx, r */
 839             o(opc | r);
 840             g(c);
 841         } else {
 842             /* we generate the shift in ecx */
 843             gv2(RC_INT, RC_ECX);
 844             r = vtop[-1].r;
 845             o(0xd3); /* shl/shr/sar %cl, r */
 846             o(opc | r);
 847         }
 848         vtop--;
 849         break;
 850     case '/':
 851     case TOK_UDIV:
 852     case TOK_PDIV:
 853     case '%':
 854     case TOK_UMOD:
 855     case TOK_UMULL:
 856         /* first operand must be in eax */
 857         /* XXX: need better constraint for second operand */
 858         gv2(RC_EAX, RC_ECX);
 859         r = vtop[-1].r;
 860         fr = vtop[0].r;
 861         vtop--;
 862         save_reg(TREG_EDX);
 863         /* save EAX too if used otherwise */
 864         save_reg_upstack(TREG_EAX, 1);
 865         if (op == TOK_UMULL) {
 866             o(0xf7); /* mul fr */
 867             o(0xe0 + fr);
 868             vtop->r2 = TREG_EDX;
 869             r = TREG_EAX;
 870         } else {
 871             if (op == TOK_UDIV || op == TOK_UMOD) {
 872                 o(0xf7d231); /* xor %edx, %edx, div fr, %eax */
 873                 o(0xf0 + fr);
 874             } else {
 875                 o(0xf799); /* cltd, idiv fr, %eax */
 876                 o(0xf8 + fr);
 877             }
 878             if (op == '%' || op == TOK_UMOD)
 879                 r = TREG_EDX;
 880             else
 881                 r = TREG_EAX;
 882         }
 883         vtop->r = r;
 884         break;
 885     default:
 886         opc = 7;
 887         goto gen_op8;
 888     }
 889 }
 890
 891 /* generate a floating point operation 'v = t1 op t2' instruction. The
 892    two operands are guaranted to have the same floating point type */
 893 /* XXX: need to use ST1 too */
 894 ST_FUNC void gen_opf(int op)
 895 {
 896     int a, ft, fc, swapped, r;
 897
 898     /* convert constants to memory references */
 899     if ((vtop[-1].r & (VT_VALMASK | VT_LVAL)) == VT_CONST) {
 900         vswap();
 901         gv(RC_FLOAT);
 902         vswap();
 903     }
 904     if ((vtop[0].r & (VT_VALMASK | VT_LVAL)) == VT_CONST)
 905         gv(RC_FLOAT);
 906
 907     /* must put at least one value in the floating point register */
 908     if ((vtop[-1].r & VT_LVAL) &&
 909         (vtop[0].r & VT_LVAL)) {
 910         vswap();
 911         gv(RC_FLOAT);
 912         vswap();
 913     }
 914     swapped = 0;
 915     /* swap the stack if needed so that t1 is the register and t2 is
 916        the memory reference */
 917     if (vtop[-1].r & VT_LVAL) {
 918         vswap();
 919         swapped = 1;
 920     }
 921     if (op >= TOK_ULT && op <= TOK_GT) {
 922         /* load on stack second operand */
 923         load(TREG_ST0, vtop);
 924         save_reg(TREG_EAX); /* eax is used by FP comparison code */
 925         if (op == TOK_GE || op == TOK_GT)
 926             swapped = !swapped;
 927         else if (op == TOK_EQ || op == TOK_NE)
 928             swapped = 0;
 929         if (swapped)
 930             o(0xc9d9); /* fxch %st(1) */
 931         if (op == TOK_EQ || op == TOK_NE)
 932             o(0xe9da); /* fucompp */
 933         else
 934             o(0xd9de); /* fcompp */
 935         o(0xe0df); /* fnstsw %ax */
 936         if (op == TOK_EQ) {
 937             o(0x45e480); /* and $0x45, %ah */
 938             o(0x40fC80); /* cmp $0x40, %ah */
 939         } else if (op == TOK_NE) {
 940             o(0x45e480); /* and $0x45, %ah */
 941             o(0x40f480); /* xor $0x40, %ah */
 942             op = TOK_NE;
 943         } else if (op == TOK_GE || op == TOK_LE) {
 944             o(0x05c4f6); /* test $0x05, %ah */
 945             op = TOK_EQ;
 946         } else {
 947             o(0x45c4f6); /* test $0x45, %ah */
 948             op = TOK_EQ;
 949         }
 950         vtop--;
 951         vtop->r = VT_CMP;
 952         vtop->c.i = op;
 953     } else {
 954         /* no memory reference possible for long double operations */
 955         if ((vtop->type.t & VT_BTYPE) == VT_LDOUBLE) {
 956             load(TREG_ST0, vtop);
 957             swapped = !swapped;
 958         }
 959
 960         switch(op) {
 961         default:
 962         case '+':
 963             a = 0;
 964             break;
 965         case '-':
 966             a = 4;
 967             if (swapped)
 968                 a++;
 969             break;
 970         case '*':
 971             a = 1;
 972             break;
 973         case '/':
 974             a = 6;
 975             if (swapped)
 976                 a++;
 977             break;
 978         }
 979         ft = vtop->type.t;
 980         fc = vtop->c.i;
 981         if ((ft & VT_BTYPE) == VT_LDOUBLE) {
 982             o(0xde); /* fxxxp %st, %st(1) */
 983             o(0xc1 + (a << 3));
 984         } else {
 985             /* if saved lvalue, then we must reload it */
 986             r = vtop->r;
 987             if ((r & VT_VALMASK) == VT_LLOCAL) {
 988                 SValue v1;
 989                 r = get_reg(RC_INT);
 990                 v1.type.t = VT_INT;
 991                 v1.r = VT_LOCAL | VT_LVAL;
 992                 v1.c.i = fc;
 993                 load(r, &v1);
 994                 fc = 0;
 995             }
 996
 997             if ((ft & VT_BTYPE) == VT_DOUBLE)
 998                 o(0xdc);
 999             else
1000                 o(0xd8);
1001             gen_modrm(a, r, vtop->sym, fc);
1002         }
1003         vtop--;
1004     }
1005 }
1006
1007 /* convert integers to fp 't' type. Must handle 'int', 'unsigned int'
1008    and 'long long' cases. */
1009 ST_FUNC void gen_cvt_itof(int t)
1010 {
1011     save_reg(TREG_ST0);
1012     gv(RC_INT);
1013     if ((vtop->type.t & VT_BTYPE) == VT_LLONG) {
1014         /* signed long long to float/double/long double (unsigned case
1015            is handled generically) */
1016         o(0x50 + vtop->r2); /* push r2 */
1017         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1018         o(0x242cdf); /* fildll (%esp) */
1019         o(0x08c483); /* add $8, %esp */
1020     } else if ((vtop->type.t & (VT_BTYPE | VT_UNSIGNED)) ==
1021                (VT_INT | VT_UNSIGNED)) {
1022         /* unsigned int to float/double/long double */
1023         o(0x6a); /* push $0 */
1024         g(0x00);
1025         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1026         o(0x242cdf); /* fildll (%esp) */
1027         o(0x08c483); /* add $8, %esp */
1028     } else {
1029         /* int to float/double/long double */
1030         o(0x50 + (vtop->r & VT_VALMASK)); /* push r */
1031         o(0x2404db); /* fildl (%esp) */
1032         o(0x04c483); /* add $4, %esp */
1033     }
1034     vtop->r = TREG_ST0;
1035 }
1036
1037 /* convert fp to int 't' type */
1038 ST_FUNC void gen_cvt_ftoi(int t)
1039 {
1040 #if 1
1041     gv(RC_FLOAT);
1042     save_reg(TREG_EAX);
1043     save_reg(TREG_EDX);
1044     gen_static_call(TOK___tcc_cvt_ftol);
1045     vtop->r = TREG_EAX; /* mark reg as used */
1046     if (t == VT_LLONG)
1047         vtop->r2 = TREG_EDX;
1048 #else
1049     int bt = vtop->type.t & VT_BTYPE;
1050     if (bt == VT_FLOAT)
1051         vpush_global_sym(&func_old_type, TOK___fixsfdi);
1052     else if (bt == VT_LDOUBLE)
1053         vpush_global_sym(&func_old_type, TOK___fixxfdi);
1054     else
1055         vpush_global_sym(&func_old_type, TOK___fixdfdi);
1056     vswap();
1057     gfunc_call(1);
1058     vpushi(0);
1059     vtop->r = REG_IRET;
1060     vtop->r2 = REG_LRET;
1061 #endif
1062 }
1063
1064 /* convert from one floating point type to another */
1065 ST_FUNC void gen_cvt_ftof(int t)
1066 {
1067     /* all we have to do on i386 is to put the float in a register */
1068     gv(RC_FLOAT);
1069 }
1070
1071 /* computed goto support */
1072 ST_FUNC void ggoto(void)
1073 {
1074     gcall_or_jmp(1);
1075     vtop--;
1076 }
1077
1078 /* bound check support functions */
1079 #ifdef CONFIG_TCC_BCHECK
1080
1081 /* generate a bounded pointer addition */
1082 ST_FUNC void gen_bounded_ptr_add(void)
1083 {
1084     /* prepare fast i386 function call (args in eax and edx) */
1085     gv2(RC_EAX, RC_EDX);
1086     /* save all temporary registers */
1087     vtop -= 2;
1088     save_regs(0);
1089     /* do a fast function call */
1090     gen_static_call(TOK___bound_ptr_add);
1091     /* returned pointer is in eax */
1092     vtop++;
1093     vtop->r = TREG_EAX | VT_BOUNDED;
1094     /* address of bounding function call point */
1095     vtop->c.i = (cur_text_section->reloc->data_offset - sizeof(Elf32_Rel));
1096 }
1097
1098 /* patch pointer addition in vtop so that pointer dereferencing is
1099    also tested */
1100 ST_FUNC void gen_bounded_ptr_deref(void)
1101 {
1102     addr_t func;
1103     int  size, align;
1104     Elf32_Rel *rel;
1105     Sym *sym;
1106
1107     size = 0;
1108     /* XXX: put that code in generic part of tcc */
1109     if (!is_float(vtop->type.t)) {
1110         if (vtop->r & VT_LVAL_BYTE)
1111             size = 1;
1112         else if (vtop->r & VT_LVAL_SHORT)
1113             size = 2;
1114     }
1115     if (!size)
1116         size = type_size(&vtop->type, &align);
1117     switch(size) {
1118     case  1: func = TOK___bound_ptr_indir1; break;
1119     case  2: func = TOK___bound_ptr_indir2; break;
1120     case  4: func = TOK___bound_ptr_indir4; break;
1121     case  8: func = TOK___bound_ptr_indir8; break;
1122     case 12: func = TOK___bound_ptr_indir12; break;
1123     case 16: func = TOK___bound_ptr_indir16; break;
1124     default:
1125         tcc_error("unhandled size when dereferencing bounded pointer");
1126         func = 0;
1127         break;
1128     }
1129
1130     /* patch relocation */
1131     /* XXX: find a better solution ? */
1132     rel = (Elf32_Rel *)(cur_text_section->reloc->data + vtop->c.i);
1133     sym = external_global_sym(func, &func_old_type, 0);
1134     if (!sym->c)
1135         put_extern_sym(sym, NULL, 0, 0);
1136     rel->r_info = ELF32_R_INFO(sym->c, ELF32_R_TYPE(rel->r_info));
1137 }
1138 #endif
1139
1140 /* Save the stack pointer onto the stack */
1141 ST_FUNC void gen_vla_sp_save(int addr) {
1142     /* mov %esp,addr(%ebp)*/
1143     o(0x89);
1144     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1145 }
1146
1147 /* Restore the SP from a location on the stack */
1148 ST_FUNC void gen_vla_sp_restore(int addr) {
1149     o(0x8b);
1150     gen_modrm(TREG_ESP, VT_LOCAL, NULL, addr);
1151 }
1152
1153 /* Subtract from the stack pointer, and push the resulting value onto the stack */
1154 ST_FUNC void gen_vla_alloc(CType *type, int align) {
1155 #ifdef TCC_TARGET_PE
1156     /* alloca does more than just adjust %rsp on Windows */
1157     vpush_global_sym(&func_old_type, TOK_alloca);
1158     vswap(); /* Move alloca ref past allocation size */
1159     gfunc_call(1);
1160 #else
1161     int r;
1162     r = gv(RC_INT); /* allocation size */
1163     /* sub r,%rsp */
1164     o(0x2b);
1165     o(0xe0 | r);
1166     /* We align to 16 bytes rather than align */
1167     /* and ~15, %esp */
1168     o(0xf0e483);
1169     vpop();
1170 #endif
1171 }
1172
1173 /* end of X86 code generator */
1174 /*************************************************************/
1175 #endif
1176 /*************************************************************/